Recommender systems survey_动视

Recommender systems survey

2025-10-03 15:21:23 责编:小OO

点击下载本文 文档为doc格式

Recommender systems survey

J.Bobadilla ⇑,F.Ortega,A.Hernando,A.Gutiérrez

Universidad Politécnica de Madrid,Ctra.De Valencia,Km.7,28031Madrid,Spain

a r t i c l e i n f o Article history:

Received 7October 2012

Received in revised form 4March 2013Accepted 19March 2013

Available online 6April 2013Keywords:

Recommender systems Collaborative ﬁltering Similarity measures Evaluation metrics Prediction

Recommendation Hybrid Social

Internet of things Cold-start

a b s t r a c t

Recommender systems have developed in parallel with the web.They were initially based on demo-graphic,content-based and collaborative ﬁltering.Currently,these systems are incorporating social infor-mation.In the future,they will use implicit,local and personal information from the Internet of things.This article provides an overview of recommender systems as well as collaborative ﬁltering methods and algorithms;it also explains their evolution,provides an original classiﬁcation for these systems,iden-tiﬁes areas of future implementation and develops certain areas selected for past,present or future importance.

1.Introduction

Recommender Systems (RSs)collect information on the prefer-ences of its users for a set of items (e.g.,movies,songs,books,jokes,gadgets,applications,websites,travel destinations and e-learning material).The information can be acquired explicitly (typically by collecting users’ratings)or implicitly [134,60,1](typically by monitoring users’behavior,such as songs heard,applications downloaded,web sites visited and books read).RS may use demo-graphic features of users (like age,nationality,gender).Social information,like followers,followed,twits,and posts,is commonly used in Web 2.0.There is a growing tend towards the use of infor-mation from Internet of things (e.g.,GPS locations,RFID,real-time health signals).

RS make use of different sources of information for providing users with predictions and recommendations of items.They try to balance factors like accuracy,novelty,dispersity and stability in the recommendations.Collaborative Filtering (CF)methods play an important role in the recommendation,although they are often used along with other ﬁlterning techniques like content-based,knowledge-based or social ones.

CF is based on the way in which humans have made decisions throughout history:besides on our own experiences,we also base our decisions on the experiences and knowledge that reach each of us from a relatively large group of acquaintances.

Recently,RS implementation in the Internet has increased,which has facilitated its use in diverse areas [171].The most com-mon research papers are focused on movie recommendation stud-ies [53,230];however,a great volume of literature for RS is centered on different topics,such as music [134,162,216],televi-sion [238,18],books [1,88],documents [206,184,183,185],e-learning [241,30],e-commerce [104,54],applications in markets [67]and web search [154],among others.

The kinds of ﬁltering most used at the beginning of the RS (col-laborative,content-based and demographic)were described in [177].Breese et al.[43]evaluated the predictive accuracy of differ-ent algorithms for CF;later,the classical paper [94]describes the base for evaluating the Collaborative Filtering RS.

The evolution of RS has shown the importance of hybrid tech-niques of RS,which merge different techniques in order to get the advantages of each of them.A survey focused on the hybrid RS has been presented in [47].However,it does not deal with the role of social-ﬁltering,a technique which has become more popular in the recent years through social networks.

The neighborhood-based CF has been the recommendation method most popular at the beginning of the RS;Herlocker et al.[93]provides a set of guidelines for designing neighborhood-based prediction systems.Adomavicius and Tuzhilin [3]present an over-view on the RS ﬁeld standing out the most complex areas on which

Corresponding author.Tel.:+34913365133;fax:+34913367527.

E-mail address:jesus.bobadilla@upm.es (J.Bobadilla).

While researchers have been developing RS,different survey papers have been published summarizing the most important is-sues in thisﬁeld.In view of the impossibility of showing every de-tail of all these techniques in just a paper,this publication selects those issues the authors have felt most suitable to understand the evolution of RS.

While the existing surveys focus on the most relevant methods and algorithms of the RSﬁeld,our survey instead tries to enhance the evolution of the RS:from aﬁrst phase based on the tradi-tional Web to the present second phase based on social Web, which is presently progressing to a third phase(Internet of things).With the purpose of being useful to the new readers of RSﬁeld,we have included in this survey some traditional topics: RS foundations,k-Nearest Neighbors algorithm,cold-start issues, similarity measures,and evaluation of RS.The rest of the paper deals with novel topics that existing surveys do not consider. Through this survey,advanced readers in RS will study in depth concepts,classiﬁcations and approaches related to social informa-tion(socialﬁltering:followers,followed,trust,reputation,credi-bility,content-basedﬁltering of social data;social tagging and taxonomies),recommending to groups of users and explaining recommendations.Readers interested in brand new and future applications willﬁnd this survey useful since it informs about the most recent works in location-aware RS trends and bio-in-spired approaches.They will also discover some important issues, such as privacy,security,P2P information and Internet of things use(RFID data,health parameters,surveillance data,teleopera-tion,telepresence,etc.).

According to the idea that RS tend to make use of different sources of information(collaborative,social,demographic,content, knowledge-based,geographic,sensors,tags,implicit and explicit data acquisition,etc.),this survey emphasizes hybrid architectures, based on making recommendations through different known tech-nologies(each one designed on behalf of a speciﬁc source of information).

Much of the quality of a survey can be measured by an appro-priate choice of its references.This survey contains249references systematically obtained,which have been selected taking into ac-count factors like the number of recent citations and the impor-tance of the journal in which the paper has been published.

The remainder of this article is structured as follows:In Sec-tion2,we explain concisely the methodology used to select the most signiﬁcative papers on the RSﬁeld.Section3describes the RS foundations:methods,algorithms and models used for provid-ing recommendations based from the information of the tradi-tional web:ratings,demographic data and item data(CF, demographicﬁltering,content-basedﬁltering and hybridﬁltering). Section4describes measures for evaluating the quality of the RS predictions and recommendations.Section5shows the use of so-cial information from Web2.0for making recomendations through concepts like trust,reputation and credibility.We will also de-scribe techniques based on content-based for social information (e.g.tags and posts).Section6focusses on two important areas (although not very well studied yet):recommendation to group of users and explanation of recommendations.Section7focusses on recommender system trends,covering bio-inspired approaches and Web3.0informationﬁltering such as location-aware RS.Sec-tion8explains related works and the original contributions of this survey.

The concluding section summarizes the RS history and focuses on the type of data used as well as the development of algorithms and evaluation measures.The conclusions section also indicates seven new areas that we consider likely to be the focus of RS re-search in the scientiﬁc community in the near future.

2.Methodology

An initial study was performed to determine the most represen-tative topics and terms in the RSﬁeld.First,300RS papers were se-lected from journals,with a higher priority for current and for often-cited articles.Next,we extracted from these300papers the most signiﬁcant terms.We gave the most emphasis to keywords, less emphasis to titles and,ﬁnally,the least emphasis to abstracts.

We have overlooked common words,like articles,prepositions and general-use words from the remaining pool,we selected300 terms represented in the RSﬁeld.From a matrix of arti-clesÂwords,wherein we stored the importance of each word from each article,we generated a tree of relationships between the words.Fig.1depicts the most signiﬁcant section of the graph (due to space constraints,the entire tree is not shown,but it is pro-vided as additional material in Fig.1AdditionalData.png).The short distances between words indicate the highest similarities;warm colors indicate a greater reliability for the relationships.The size of the nodes indicates the importance of the words as a function of the parameters N k,N t,N a(number of signiﬁcative words in the keywords,title and abstract)and N k

;N t

;N a

(number of times that the word w appears in the keywords,title and abstract).The equa-tion used to determine the importance of each word w is as follows:

f w¼

1N k

N t

N log N a

N t

N a

N N a

N t

Example:we will consider a paper where N k=5keywords,N t=11 words in the title,and N a=52words of abstract length.We will get the values of f factorization and f matrix,where the word‘factorization’appears once as a keyword,once in the title and three times in the abstract;the word‘matrix’does not appear as a keyword,but it is contained once in the title and twice in the abstract.The importance of these words will be:

f factorization¼

11log52

5252

¼0:09

matrix

11log52

5252

¼0:02

The information depicted in Fig.1is used to identify the most relevant aspects of RS.They are represented by the most signiﬁcant words in the graph and the related terms.The articles referenced herein were chosen based on the following criteria:(a)the tran-scendence of the subject according to the importance of the words in Fig.1;(b)its historical contribution(a signiﬁcant fraction of the classic reference articles are included);(c)the number of times the article is cited;(d)articles published in journals with an impact factor were preferred over conferences and workshops;and(e)re-cent articles were preferred over articles published many years ago.Fig.2shows a temporal distribution for the referenced papers.

We use the clusters of words in Fig.1to structure the explica-tions of the survey.For each concept explained:(1)we have ob-tained their keywords and all the words related to them according to Fig.1;(2)we have identiﬁed,among the set of300pa-pers,those which are more related to the set of words associated to the concept;(3)we have selected the subset of papers which deal with the concept,giving priority to those with high values in crite-ria like importance of the paper and the number of cites;and(4) we have tried to balance the number of times a paper is referenced in our survey,aiming to reference most of the300papers selected.

110J.Bobadilla et al./Knowledge-Based Systems46(2013)109–1323.Recommender systems foundations

This section presents the most relevant concepts on which the traditional RS are based.Here,we provide general descriptions on the classical taxonomies,algorithms,methods,ﬁltering ap-proaches,databases,etc.Besides,we show a graphic depicting the traditional models of recommendations and their relations. Next,we will describe the cold-start problem,which will illustrate the difﬁculty of making collaborative recommendation when the RS contains a small amount of data.Next,we will describe the k NN algorithm;the most used algorithm for implementing RS based on CF.Finally,we will describe different proposed similarity measures for comparing users or items.We will show graphics for measuring the quality of these similarity measures.

3.1.Fundamentals

The process for generating an RS recommendation is based on a combination of the following considerations:

The type of data available in its database(e.g.,

istration information,features and content for

ranked,social relationships among users and

information).

Theﬁltering algorithm used(e.g.,demographic,

collaborative,social-based,context-aware and

The model chosen(e.g.,based on direct use of

based,’’or a model generated using such data:

The employed techniques are also considered:

approaches,Bayesian networks,nearest neighbors

bio-inspired algorithms such as neural networks

algorithms;fuzzy models,singular value decomposition

niques to reduce sparsity levels,etc.

Sparsity level of the database and the desired

Performance of the system(time and memory consuming).

The objective sought is considered(e.g.,predictions and top N recommendations)as well as

The desired quality of the results(e.g.,novelty,coverage and precision).

Research in RS requires using a representative set of public dat-abases to facilitate investigations on the techniques,methods and algorithms developed by researchers in theﬁeld.Through these databases,the scientiﬁc community can replicate experiments to validate and improve their techniques.Table1lists the current public databases referenced most often in the literature.Last.Fm and Delicious incorporate implicit ratings and social information; their data were generated from the versions released in the HetRec, 2011data sets,hosted by the GroupLens research Group.

The internal functions for RS are characterized by theﬁltering algorithm.The most widely used classiﬁcation divides theﬁltering algorithms into[3,51,203]:(a)collaborativeﬁltering,(b)demo-graphicﬁltering,(c)content-basedﬁltering and(d)hybridﬁltering.

represented in the recommender systems researchﬁeld.Short distances indicate higher similarities,and a warm color indicates greater proportional to the importance of the words.

Fig.2.Temporal distribution for the referenced papers.112J.Bobadilla et al./Knowledge-Based Systems46(2013)109–132

Table1

Most often used memory-based recommender systems public databases.

Without social information With social information(hosted by the GroupLens)

MovieLens1M MovieLens10M Netﬂix Jester EachMovie Book-crossing ML Last.Fm Delicious

Ratings1million10million100million 4.1million 2.8million 1.1million855,592,834104,833 Users604071,567480,173,42172,916278,8582113121867 Items359210,68117,7701001628271,37910,15317,63269,226 Range{1,...,5}{1,...,5}{1,...,5}À10,10[0,1]{1,...,10}{1,...,5}Implicit Implicit Tags N/A N/A N/A N/A N/A N/A132221194653388 Tags assignment N/A N/A N/A N/A N/A N/A479571879437593 Friends relations N/A N/A N/A N/A N/A N/A N/A2543415328

Fig.3.Traditional models of recommendations and their relationships.implicitly acquired(e.g.,number of times a song is heard,informa-tion consulted and access to a resource).

The most widely used algorithm for collaborativeﬁltering is the k Nearest Neighbors(kNN)[3,203,32].In the user to user version, k NN executes the following three tasks to generate recommenda-tions for an active user:(1)determine k users neighbors(neighbor-hood)for the active user a;(2)implement an aggregation approach with the ratings for the neighborhood in items not rated by a;and (3)extract the predictions from in step2then select the top N recommendations.

Hybridﬁltering[47,185].Commonly uses a combination of CF with demographicﬁltering[224]or CF with content-basedﬁltering [18,60]to exploit merits of each one of these techniques.Hybrid ﬁltering is usually based on bioinspired or probabilistic methods such as genetic algorithms[76,99],fuzzy genetic[7],neural net-works[133,62,192],Bayesian networks[50],clustering[209]and latent features[199].

A widely accepted taxonomy divides recommendation methods into memory-based and model-based method categories: Memory-based methods[3,51,123,214].Memory-based methods can be deﬁned as methods that(a)act only on the matrix of user ratings for items and(b)use any rating generated before the refer-ral process(i.e.,its results are always updated).Memory-based methods usually use similarity metrics to obtain the distance be-tween two users,or two items,based on each of their ratios.

Model-based methods[3,212].Use RS information to create a model that generates the recommendations.Herein,we consider a method model-based if new information from any user outdates the model.Among the most widely used models we have Bayesian classiﬁers[59],neural networks[107],fuzzy systems[234],genetic algorithms[76,99],latent features[251]and matrix factorization [142],among others.

To reduce the problems from high levels of sparsity in RS dat-abases,certain studies have used dimensionality reduction tech-niques[202].The reduction methods are based on Matrix Factorization[124,142,143].Matrix factorization is especially ade-quate for processing large RS databases and providing scalable ap-proaches[215].The model-based technique Latent Semantic Index (LSI)and the reduction method Singular Value Decomposition (SVD)are typically combined[224,244,48].SVD methods provide good prediction results but are computationally very expensive; they can only be deployed in static off-line settings where the known preference information does not change with time.

RS can use clustering techniques to improve the prediction qual-ity and reduce the cold-start problem when applied to hybridﬁl-tering.It is typical to form clusters of items in hybrid RS [209,237].A different common approach uses clustering both for items and users(bi-clustering)[252,85].RS comprising social infor-mation have been clustered to improve the following areas:tagging [208],explicit social links[179]and explicit trust information [181,70].

The graph in Fig.3shows the most signiﬁcant traditional meth-ods,techniques and algorithms for the recommendation process as well as their relationships and groupings.Different sections of this paper provide more detail on the most important aspects involved in the recommendation process.

As may be seen in Fig.3,we can use some of the traditionalﬁl-tering methods(content-based,demographic and collaborative) applied to databases.Model-based technologies(genetic algo-rithms,neural networks,etc.)make use of this kind of information. Typical memory-based approaches are:item to item;user to user; and hybrids of the two previous.The main purpose of both mem-ory-based and model-based approaches is to get the most accurate predictions in the tastes of users.The accuracy of these predictions may be evaluated through the classical information retrieval mea-sures,like MAE,precision,and recall.Researchers make use of these measures in order to improve the RS methods and technologies.

3.2.Cold-start

The cold-start problem[203,3]occurs when it is not possible to make reliable recommendations due to an initial lack of ratings. We can distinguish three kinds of cold-start problems:new com-munity,new item and new user.The last kind is the most important in RS that are already in operation.

The new community problem[204,129]refers to the difﬁculty, when starting up a RS,in obtaining,a sufﬁcient amount of data (ratings)for making reliable recommendations.Two common ways are used for tackling this problem:to encourage users to make ratings through different means;to take CF-based recom-mendations when there are enough users and ratings.

The new item problem[174,172]arises because the new items entered in RS do not usually have initial ratings,and therefore,they are not likely to be recommended.In turn,an item that is not rec-ommended goes unnoticed by a large part of the community of users,and as they are unaware of it they do not rate it;this way, we can enter a vicious circle in which a set of items of the RS are left out of the ratings/recommendations process.The new item problem has less of an impact on RS in which the items can be dis-covered via other means(e.g.movies)than in RS where this is not the case(i.e.e-commerce,blogs,photos,videos,etc.).A common solution to this problem is to have a set of motivated users who are responsible for rating each new item in the system.

The new user problem[190,197]represents one of the great dif-ﬁculties faced by the RS in operation.Since new users in the RS have not yet provided any rating in the RS,they cannot receive any personalized recommendations based on memory-based CF; when the users enter theirﬁrsts ratings they expect the RS to offer them personalized recommendations,but the number of ratings introduced in the RS is usually not yet sufﬁcient to be able to make reliable CF-based recommendations,and,therefore,new users may feel that the RS does not offer the service they expected and they may stop using it.

The common strategy to tackle the new user problem consists of turning to additional information to the set of ratings in order to be able to make recommendations based on the data available for each user.The cold-start problem is often faced using hybrid approaches(usually CF-content based RS,CF-demographic based RS,CF-social based RS)[118,140].Leung et al.[135]propose a no-vel content-based hybrid approach that makes use of cross-level association rules to integrate content information about domains items.Kim et al.[118]use collaborative tagging employed as an approach in order to grasp andﬁlter users’preferences for items and they explore the advantages of the collaborative tagging for data sparseness and cold-start users(they collected the dataset by crawling the collaborative tagging delicious site).Weng et al. [228]combine the implicit relations between users’items prefer-ences and the additional taxonomic preferences to make better quality recommendations as well as alleviate the cold-start prob-lem.Loh et al.[140]represent user’s proﬁles with information ex-tracted from their scientiﬁc publications.Martinez et al.[148] present a hybrid RS which combines a CF algorithm with a knowl-edge-based one.Chen and He[56]propose a number of common terms/term frequency(NCT/TF)CF algorithm based on demo-graphic vector.Saranya and Atsuhiro[199]propose a hybrid RS that utilizes latent features extracted from items represented by a multi-attributed record using a probabilistic model.Park et al. [173]propose a new approach:they useﬁlterbots,and surrogate users that rate items based only on user or item attributes.

J.Bobadilla et al./Knowledge-Based Systems46(2013)109–132113

3.3.The k nearest neighbors recommendation algorithm

The k Nearest Neighbors (k NN)recommendation algorithm is the reference algorithm for the collaborative ﬁltering recommendation process.Its primary virtues are simplicity and reasonably accurate results;its major pitfalls are low scalability and vulnerability to sparsity in the RS databases.This section provides a general expla-nation of this algorithm function.

CF based on the k NN algorithm is conceptually simple,with a straightforward implementation;it also generally produces good-quality predictions and recommendations.However,due to the high level of sparsity [142,29]in RS databases,similarity measures often encounter processing problems (typically from insufﬁcient mutual ratings for a comparison of users and items)and cold start situations (users and items with low number of rankings)[204,98,36,135].

Another major problem for the k NN algorithm is its low scalabil-ity [142].As the databases (such as Netﬂix)increase in size (hun-dreds of thousands of users,tens of thousands of items,and hundreds of millions of rankings),the process for generating a neighborhood for an active user becomes too slow;The similarity measure must be processed as often as new users are registered in the database.The item to item version of the k NN algorithm sig-niﬁcantly reduces the scalability problem [200].To this end,neigh-bors are calculated for each item;their top n similarity values are stored,and for a period of time,predictions and recommendations are generated using the stored information.Although the stored information does not include the ratings from previous process-ing/storage,outdated information for items is less sensitive than for the users.

A recurrent theme in CF research is generating metrics to calcu-late with accuracy and precision the existing similarity for the users (or items).Traditionally,a series of statistical metrics have been used [3,51],such as the Pearson correlation ,cosine ,constraint Pearson correlation and mean squared differences .Recently,metrics have been designed to ﬁt the constraints and peculiarities of RS [31,35].The relevance (signiﬁcance )concept was introduced to af-ford more importance to more relevant users and items [34,227].Additionally,a group of metrics was speciﬁcally designed to ade-quately function in cold-start situations [6,36].

The k NN algorithm is based on similarity measures.Next sub-section provides further details on the current RS similarity mea-sures.The similarity approaches typically compute the similarity between two users x and y (user to user)based on both users’item ratings.The item to item k NN version computes the similarity be-tween two items i and j .

A formal approach of the k NN algorithm may be found in [32].In this section,we will provide an illustrative example of this algo-rithm.The method for making recommendations is based on the following three steps:

(a)Using the selected similarity measure,we produce the set of

k neighbors for the active user a .The k neighbors for a are the nearest k (similar)users to u .

(b)Once the set of k users (neighbors)similar to active a has

been calculated,in order to obtain the prediction of item i on user a ,one of the following aggregation approaches is often used:the average,the weighted sum and the adjusted weighted aggregation (deviation-from-mean).

(c)To obtain the top-n recommendations,we choose the n

items,which provide most satisfaction to the active user according to our predictions.Fig.4shows a case study using the user to user k NN algorithm mechanism.

In the item to item version [200,77]of the k NN algorithm,the following three tasks are executed:(1)determine q items neigh-bors for each item in the database;(2)for each item i not ranked by the active user a ,calculate its prediction based on the ratings of a from the q neighbors of i ;and (3)select the top n recommen-dations for the active user (typically the n major predictions from a ).Step (1)can be executed periodically,which facilitates an accel-erated recommendation with regard to the user to user version.The item to item and user to user versions of the k NN algorithm can be combined [188]to take advantage of the positive aspects from each approach.These approaches are typically fused by pro-cessing the similarity between objects.3.4.Similarity measures

A metric or a Similarity Measure (SM)determines the similarity between pairs of users (user to user CF)or the similarity between pairs of items (item to item CF).For this purpose,we compare the ratings of all the items rated by two users (user to user)or the rat-ings of all users who have rated two items (item to item).

The k NN algorithm is based essentially on the use of traditional similarity metrics of statistical origin.These metrics require,as the only source of information,the set of votes made by the users on the items (memory-based CF).Among the most commonly used traditional metrics we have:Pearson correlation (CORR),cosine (COS),adjusted cosine (ACOS),constrained correlation (CCORR),Mean Squared Differences (MSD)and Euclidean (EUC)[51,3].

We will describe and compare a representative group of SM used in the k NN algorithm.The SM discussed include the following variations:(a)cold-start and general cases,(b)based or not based on models,and (c)using trust information or only ratings.Table 2shows a classiﬁcation of the memory-based CF SM which will be tested in this section.

A new metric (JMSD)has recently been published,which be-sides using the numerical information from the ratings (via mean squared differences)also uses the non-numerical information pro-vided by the arrangement of these (via Jaccard)[31].Ortega et al.[169]use Pareto dominance to perform a pre-ﬁltering process eliminating less representative users from the k -neighbur selection process while retaining the most promising ones.

A specialization of the memory-based CF SM,which appeared recently [35],uses the information contained in the votes of all users,instead of restricting it to the ratings of the two users com-pared (user to user)or the two items compared (item to item).We will call this SM SING (singularities).

The possibility exists to create a model (model-based CF)from the full set of users’ratings in order to later determine the similar-ity between pairs of users or pairs of items based on the model cre-ated.The potential advantages of this focus are an increase in the accuracy obtained,in the performance (time consuming)achieved or in both.The drawback is that the model must be regularly up-dated in order to consider the most recently entered set of

ratings.

Fig.4.User to user k NN algorithm example,k =3.Similarity measure:1–(mean squared differences).Aggregation approach:average.

Systems 46(2013)109–132

Bobadilla et al.[33]provides a metric based on a model generated using genetic algorithms.We will call this SM GEN(genetic-based).

As a result of the increase in web2.0websites on the Internet,a set of metrics has appeared which use the new social information available(friends,followers,followeds,etc.).Most of these SM are grouped in papers related to trust,reputation and credibility [71,239,138],although this situation is also produced in other ﬁelds[30].These metrics could not be considered strictly mem-ory-based CF,as they use additional information which not all RS have.In this sense,each SM proposed is tailored to a speciﬁc RS or at most to a very small set of RS which share the same structure in their social information.

There are SM[112,127]which aim to extract information re-lated to trust and reputation by only using the users’set of ratings (memory-based CF).The advantage is that their use can be general-ized to all CF RS;the drawback is that the social information ex-tracted is really poor.We will call TRUST the SM proponed in Jeong et al.[112].Currently,two new interesting SM get more cov-erage[38]and accuracy[61].

Fig.5shows the results from several evaluation measures gen-erated by applying the SM discussed in this section.The results show that the RS-tailored SM are superior compared with the tra-ditional SM from statistics.Processing for the memory-based infor-mation and results from Fig.5follow the framework schematic published previously[32].

There are so far research papers dealing with the cold-start prob-lem through the users’ratings information.Ahn[6]presents a heu-ristic SM named PIP,that outperforms the traditional statistical SM

5.Evaluation measures results obtained from current similarities measures;MovieLens database.(A)Prediction results,(B)recommendation results,(C)novelty results,

(D)trust results.(Pearson correlation,cosine,etc.).Heung-Nam et al.[98]proposes a method(UERROR)that predictsﬁrst actual ratings and subsequently identiﬁes prediction errors for each user.Taking into account this er-ror information,some speciﬁc‘‘error-reﬂected’’models,are de-signed.Bobadilla et al.[36]presents a metric based on neural learning(model-based CF)and adapted for new user cold-start situ-ations,called NCS.

Fig.6shows results from several evaluation measures gener-ated by applying the cold-start SM presented in this section;These results show that the RS-tailored SM are superior compared with the traditional SM from statistics.Since the database Movielens does not take into account cold-start users,we have removed rat-ings of this database in order to achieve cold-start users.Indeed, we have removed randomly between5and20ratings of those users who have rated between20and30items.In this way,we will regard those users who now result to rate between2and20 items as cold-start users.

4.Evaluation of recommender systems results

Since RS research began,evaluation of predictions and recom-mendations has become important[94,201].Research in the RS ﬁeld requires quality measures and evaluation metrics[90]to know the quality of the techniques,methods,and algorithms for predic-tions and recommendations.Evaluation metrics[94,95]and evalua-tion frameworks[92,32]facilitate comparisons of several solutions for the same problem and selection from different promising lines of research that generate better results.

Because of evaluation measures,RS recommendations have gradually been tested and improved[48].A representative set of existing evaluation measures has standard formulations,and a group of open RS public databases has been generated.These two advances have facilitated quality comparisons for new proposed recommendation methods and previously published methods;thus, RS methods and algorithms research has progressed continuously.

The most commonly used quality measures are the following [90,95]:(1)prediction evaluations,(2)evaluations for recommen-dation as sets,and(3)evaluations for recommendations as ranked lists.Fig.5shows results from applying several evaluation mea-sures to a set of representative similarity measures.

Evaluation metrics[12]can be classiﬁed as[94,95](a)predic-tion metrics:such as the accuracy ones:Mean Absolute Error (MAE),Root of Mean Square Error(RMSE),Normalized Mean Average Error(NMAE);and the coverage(b)set recommendation metrics: such as Precision,Recall and Receiver Operating Characteristic (ROC)[204](c)rank recommendation metrics:such as the half-life [43]and the discounted cumulative gain[17]and(d)diversity met-rics:such as the diversity and the novelty of the recommended items[105].The validation process is performed by employing the most common cross validation techniques(random sub-sam-pling and k-fold cross validation)[21];for cold-start situations, due to the limited number of users(or items)votes involved,the usual method chosen to carry out the experiments is leave-one-out cross validation[36].

Hernández and Gaudioso[95]propose an evaluation process based on the distinction between interactive and

non-interactive 6.Evaluation results obtained from current cold-start similarities measures.(A)Prediction results,(B)recommendation results,(C)novelty results,and(D)trust results.subsystems.General publications and reviews also exist which in-clude the most commonly accepted evaluation measures:mean absolute error,coverage,precision,recall and derivatives of these: mean squared error,normalized mean absolute error,ROC and fallout; Goldberg et al.[87]focuses on the aspects not related to the eval-uation,Breese et al.[43]compare the predictive accuracy of vari-ous methods in a set of representative problem domains.

The majority of articles discuss attempted improvements to the accuracy of RS results(RMSE,MAE,etc.).It is also common to at-tempt an improvement in recommendations(precision,recall, ROC,etc.).However,additional objectives should be considered for generating greater user satisfaction[253],such as topic diversi-ﬁcation and coverage serendipity.

Currently,theﬁeld has a growing interest in generating algo-rithms with diverse and innovative recommendations,even at the expense of accuracy and precision.To evaluate these aspects, various metrics have been proposed to measure recommendation novelty and diversity[105,220].

The frameworks aid in deﬁning and standardizing the methods and algorithms employed by RS as well as the mechanisms to eval-uate the quality of the results.Among the most signiﬁcant papers that propose CF frameworks are Herlocker et al.[92]which evaluates the following:similarity weight,signiﬁcance weighting, variance weighting,selecting neighborhood and rating normaliza-tion;Hernández and Gaudioso[95]proposes a framework in which any RS is formed by two different subsystems,one of them to guide the user and the other to provide useful/interesting items. Koutrika et al.[125]is a framework which introduces levels of abstraction in CF process,making the modiﬁcations in the RS more ﬂexible.Antunes et al.[12]presents an evaluation framework assuming that evaluation is an evolving process during the system lifecicle.

The majority of RS evaluation frameworks proposed until now present two deﬁciencies:theﬁrst of these is the lack of formal-ization.Although the evaluation metrics are well deﬁned,there are a variety of details in the implementation of the methods which,in the event they are not speciﬁed,can lead to the generation of different results in similar experiments.The second deﬁciency is the absence of standardization of the evalu-ation measures in aspects such as novelty and trust of the recommendations.

Bobadilla et al.[32]provides a complete series of mathematical formalizations based on sets theory.Authors provide a set of eval-uation measures,which include the quality analysis of the follow-ing aspects:predictions,recommendations,novelty and trust.

Presented next is a representative selection of the RS evaluation quality measures most often used in the bibliography.

4.1.Quality of the predictions:mean absolute error,accuracy and coverage

In order to measure the accuracy of the results of an RS,it is usual to use the calculation of some of the most common predic-tion error metrics,amongst which the Mean Absolute Error (MAE)and its related metrics:mean squared error,root mean squared error,and normalized mean absolute error stand out.

We deﬁne U as the set of the RS users,I as the set of the RS items,r u,i the rating of user u on item i, the lack of rating(r u,i= means user u has not rated item i),p u,i the prediction of item i on user u.

Let O u={i2I j p u,i– ^r u,i– },set of items rated by user u hav-ing prediction values.We deﬁne the MAE and RMSE of the system as the average of the user’s MAE.We remark that the absolute dif-ference between prediction and real value,j p u,iÀr u,i j,informs about the error in the prediction.MAE¼

u2U

#O u

i2O u

j p u;iÀr u;i j

ð1ÞRMSE¼

u2U

ﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ

#O u

i2O u

ðp u;iÀr u;iÞ2

ð2Þ

The coverage could be deﬁned as the capacity of predicting from a metric applied to a speciﬁc RS.In short,it calculates the percent-age of situations in which at least one k-neighbor of each active user can rate an item that has not been rated yet by that active user.We deﬁned K u,i as the set of neighbors of u which have rated the item i.We deﬁne the coverage of the system as the average of the user’s coverage:

Let

C u¼f i2I j r u;i¼ ^K u;i–£g;

D u¼f i2I j r u;i¼ g

co v erage¼1#U

u2U

100Â

#C u

#D u

ð3Þ4.2.Quality of the set of recommendations:precision,recall and F1

The conﬁdence of users for a certain RS does not depend directly on the accuracy for the set of possible predictions.A user gains conﬁdence on the RS when this user agrees with a reduced set of recommendations made by the RS.

In this section,we deﬁne the following three most widely used recommendation quality measures:(1)precision,which indicates the proportion of relevant recommended items from the total number of recommended items,(2)recall,which indicates the pro-portion of relevant recommended items from the number of rele-vant items,and(3)F1,which is a combination of precision and recall.

Let X u as the set of recommendations to user u,and Z u as the set of n recommendations to user u.We will represent the evaluation precision,recall and F1measures for recommendations obtained by making n test recommendations to the user u,taking a h rele-vancy threshold.Assuming that all users accept n test recommendations:

precision¼

u2U

#f i2Z u j r u;i P h g

ð4Þrecall¼

u2U

#f i2Z u j r u;i P h g

#f i2Z u j r u;i P h gþ#i2Z c

r u;i P h

ÈÉð5ÞF1¼

2ÂprecisionÂrecall

precisionþrecall

ð6Þ4.3.Quality of the list of recommendations:rank measures

When the number n of recommended items is not small,users give greater importance to theﬁrst items on the list of recommen-dations.The mistakes incurred in these items are more serious er-rors than those in the last items on the list.The ranking measures consider this situation.Among the ranking measures most often used are the following standard information retrieval measures: (a)half-life(7)[43],which assumes an exponential decrease in the interest of users as they move away from the recommenda-tions at the top and(b)discounted cumulative gain(8)[17],wherein decay is logarithmic.

HL¼

u2U

X N

i¼1

maxðr u;p

Àd;0Þ

2ðiÀ1Þ=ðaÀ1Þ

ð7ÞDCG k¼

u2U

r u;p

X k

i¼2

r u;p

log

ðiÞ

ð8Þ

p 1,...,p n represents the recommendation list,r u ,pi represents the true rating of the user u for the item p i ,k is the rank of the eval-uated item,d is the default rating,a is the number of the item on the list such that there is a 50%chance the user will review that item.

4.4.Novelty and diversity

The novelty evaluation measure indicates the degree of differ-ence between the items recommended to and known by the user.The diversity quality measure indicates the degree of differentia-tion among recommended items.

Currently,novelty and diversity measures do not have a stan-dard;therefore,different authors propose different metrics [163,220].Certain authors have [105]used the following:

di v ersity Z u ¼1

u Z u À1ÞX i 2Z u

X j 2Z u

;j –i

½1Àsim ði ;j Þ

ð9Þno v elty i ¼

u X

j 2Z

½1Àsim ði ;j Þ ;

i 2Z u

ð10Þ

Here,sim (i ,j )indicates item to item memory-based CF similar-ity measures.Z u indicates the set of n recommendations to user u .4.5.Stability

The stability in the predictions and recommendations inﬂu-ences on the users’trust towards the RS.A RS is stable if the pre-dicitions it provides do not change strongly over a short period of time.Adomavicius and Zhang [4]propose a quality measure of stability,MAS (Mean Absolute Shift).This measure is deﬁned through a set of known ratings R 1and a set of predictions of all un-known ratings,P 1.For an interval of time,users of the RS will have rated a subset S of these unknown ratings and the RS can now make new predictions,P 2.MAS is deﬁned as follows:

stability ¼MAS ¼

j P 2j X

ðu ;i Þ2P

j P 2ðu ;i ÞÀP 1ðu ;i Þj

ð11Þ

4.6.Reliability

The reliability of a prediction or a recommendation informs about how seriously we may consider this prediction.When RS recommends an item to a user with prediction 4.5in a scale {1,...,5},this user hopes to be satisﬁed by this item.However,this value of prediction (4.5over 5)does not reﬂect with which certain degree the RS has concluded that the user will like this item (with value 4.5over 5).Indeed,this prediction of 4.5is much more reli-able if it has obtained by means of 200similar users than if it has obtained by only two similar users.

In Hernando et al.[96],a realibility measure is proposed accord-ing the usual notion that the more reliable a prediction,the less lia-ble to be wrong.Although this reliability measure is not a quality measure used for comparing different techniques of RS through cross validation,this can be regarded as a quality measure associ-ated to a prediction and a recommendation.In this way,the RS pro-vides a pair of values (prediction value,reliability value),through which users may balance its preference:for example users would probably prefer the option (4,0.9)to the option (4.5,0.1).Conse-quently,the reliability measure proposed in Hernando et al.[96]provides a new understandable factor,which users may consider for taking its decisions.Nevertheless,the use of this reliability measure is just constrained to those RS based on the k NN algorithm.

The deﬁnition of reliability on the prediction,p u ,i ,is based on two numeric factors:s u ,i and v u ,i .s u ,i measures the similar-ity of the neighbors used for making the prediction p u ,i ;v u ,i measures the degree of disagreement between these neighbors rating the item i .Finally,the reliablity measure is deﬁned as follows:

f S ðs u ;i Þ¼1À

s þs u ;i

;

s u ;i ¼

v 2K u ;i

sim ðu ;v Þ

ð12Þ

where

f S ðs u ;i Þ¼1À

s þs u ;i

;

s u ;i ¼

v 2K u ;i

sim ðu ;v Þ

ð13

Fig.7.Recommender systems evaluation process.

f v ðv u ;i Þ¼max Àmin Àv u ;i max Àmin ln 0:5ln max

Àmin

À v ;v u ;i

v 2K u ;i sim ðu ;v Þðr v ;i À

r v Àp u ;i þ r u Þ2P v 2K u ;i v ð14Þ

where s and v

are respectively the median of the values of s u ,i and v u ,i in the speciﬁc RS.K u ,i is the set of neighbors of u which have rated the item i .{min,...,max}is the discrete range of rating values.Fig.7shows the general mechanism for cross validation used to generate quality results form the evaluation measures.The data-base is divided in training and test areas for both users and items.In the ﬁrst phase (top on the left side),k -neighbors are calculated for the active user (while the active user is selected from the set of test users,the k -neighbors are selected from the set of training users).In the aggregation phase (top on the right side),predictions are calculated for the active user (from the set of test items).Final-ly,evaluation metrics are used to compare the predictions and rec-ommendations obtained with the real ratings of the user;the more accurate the predictions and recommendations,better quality of the proposed recommendation algorithm.

5.Social information

As the web 2.0has developed,RS have increasingly incorpo-rated social information (e.g.,trusted and untrusted users,fol-lowed and followers,friends lists,posts,blogs,and tags).This new contextual information [145,216]improves the RS.Social information improves the sparsity problem inherent in memory-based RS because social information reinforces traditional mem-ory-based information (users ratings):users connected by a net-work of trust exhibit signiﬁcantly higher similarity on items and meta-data that non-connected users [132].

Social information is used by researchers with three primary objectives:(a)to improve the quality of predictions and recom-mendations [53,13],(b)propose or generate new RS [139,210],and (c)elucidate the most signiﬁcant relationships between social information and collaborative processes [100,178].

Trust and reputation is an important area of research in RS [166];this area is closely related to the social information currently in-cluded in RS [114].The most common approachs to generating trust and reputation measurements are the following:(a)user trust:to calculate the credibility of users through explicit informa-tion of the rest of users [239,138]or to calculate the credibility of users through implicit information obtained in a social network [59,150]and (b)item trust:to calculate the reputation of items through a feedback of users [114]or to calculate the reputation of items studying how users work with these items [58,122].

In the social RS ﬁeld,users can introduce labels associated with items.The set of triples h user,item,tag i form information spaces referred to as folksonomies .Fundamentally,folksonomies are used in the following two ways:(1)to create tag recommendation sys-tems (RS based only on tags)[147]and (2)to enrich the recom-mendation processes using tags [81].

Content-based ﬁltering has recently become more important due to the surge in social networks.RS show a clear trend to allow users to introduce content [13,178],such as comments,critiques,ratings,opinions and labels as well as to establish social relation-ship links (e.g.,followed,followers,like user and dislike user).This additional information increases the accuracy of predictions and recommendations,which has generated a variety of research arti-cles:Kim et al.[117],Zheng and Li [248]and Carrer-Neto et al.[53].The rest of this section deal is dealt with the concepts and re-search in the two lines considered previously:Filtering of social information and content ﬁltering.

5.1.Social Filtering

Social information can be gathered explicitly or implicitly through identiﬁcation of a community network or afﬁnity network [196]using the individual information that users generate (e.g.,communications and web logs)[178].Even using only the ratings from the users,it is possible to improve the RS results creating an implicit social networking [180].Both implicit and explicit information sources can be combined to generate recommenda-tions [144].

The explicit social information can be used via a trust-based CF in order to improve the quality of recommendations.Trust infor-mation can be generated or used through different approaches,such as trust propagation mechanisms [42],a ‘follow the leader’approach [8,186],personality-based similarity measures [101],trust networks [239,221],distrust analysis [223,20],and dynamic trust based on the ant colonies metaphor [20].

Most of the research work that uses social information applied to RS aims to obtain improvements in the recommendations made by referring to the extra information provided by the social infor-mation used.Among the most relevant current work which uses this approach we have:Woerndl and Groh [231]use social net-works to enhance collaborative ﬁltering;Their evaluation shows that the social recommender outperforms traditional collaborative ﬁltering algorithms in the used scenario.Arazy et al.[13]improve accuracy by using data from online social networks and electronic communication tools.Xin et al.[233]propose an approach for improving RS through exploiting the learners note taking activity.They maintain that notes’features can be exploited by collabora-tive learning systems in order to enrich and extend the user proﬁle and improve personalized learning.The Bonhard and Sasse [41]re-search has shown that the relationship between advice-seeker and recommender is extremely important,so ways of indicating social closeness and taste overlap are required.They thus suggest that drawing on similarity and familiarity between the user and the persons who have rated the items can aid judgment and decision making.Fengkun and Hong [75]developed a way to increase rec-ommendation effectiveness by incorporating social network infor-mation into CF.They collected data about users’preference ratings and their social network relationships from a social networking web site;then,they evaluated CF performance with diverse neigh-bor groups combining groups of friends and nearest neighbors.Carmagnola et al.[52]state that joining in a network with other people exposes individuals to social dynamics which can inﬂuence their attitudes,behaviors and preferences:They present SoNARS ,an algorithm for recommending content in social RS.SoNARS tar-gets users as members of social networks,suggesting items that re-ﬂect the trend of the network itself,based on its structure and on the inﬂuence relationships among users.In Ramaswamy et al.[1]the design of the social network based RS incorporates three features that complement each other to derive highly targeted ads.First,they analyze information such as customer’s address books to estimate the level of social afﬁnity among various users.This social afﬁnity information is used to identify the recommendations to be sent to an individual user.

Another group of research work uses social information to cre-ate or enable RS.That is,the aim is not to improve the results of a particular RS in operation,the aim is to propose or make possible RS which still do not exist,or if they do exist they are not based on social information:The Siersdorfer and Sergei [210]objective is to construct social recommender systems that predict the utility of items,users,or groups based on the multi-dimensional social environment of a given user;they do a mining of the rich set of structures and social relationships that provides the folksonomies.In the Li and Chen [137]study they propose a blog recommenda-tion mechanism that combines trust model,social relation and

A third group of work provides the foundation of the research to discover the most signiﬁcant relationships between social informa-tion and collaborative processes,without creating,proposing or improving any particular RS.This research moves at a higher level of abstraction,with the aim of establishing bases and general prin-ciples.Bonhard[40]paper explains that qualitative research con-ducted to date has shown that the relationship between recommender and recommendee has a signiﬁcant impact on deci-sion-making.Hossain and Fazio[100]present a study exploring the connection between social networks and collaborative process. They focus on exploring academics’network position and its effect on their collaborative networks.By deﬁning network position in this way,they develop a social network that uses the academics as nodes within the network instead of each published paper. The Esslimani et al.[72]paper presents a new CF approach based on a behavioral network that uses navigational patterns to model relationships between users and exploits social networks tech-niques.Golbeck and Kuter[86]present an experimental study of several types of trust inference algorithms to answer the following questions on trust and change:How far does a single change prop-agate through the network?How large is the impact of that change?How does this relate to the type of inference algorithm? The experimental results provide insights into which algorithms are most suitable for certain applications.

Research in theﬁeld of trust and reputation could provide a suitable starting point to create social interaction among users of the RS,however,the most relevant work on the subject is limited to the use of trust relationships to improve the quality of the rec-ommendation services.O’donovan[165]book chapter examines the diversity of sources from which trust information can be har-nessed within social web applications and discusses a high level classiﬁcation of those sources.It is shown that harnessing an in-creased amount of information upon which to make trust decisions greatly enhances the user experience with the social web applica-tion.Massa and Avesani[151]explain that RS making use of trust information are the most effective in term of accuracy while pre-serving a good coverage.This is especially evident on users who provided few ratings.Yuan et al.[239]choose the trust aware RS as an example to demonstrate the advantages by making use of the veriﬁed small-world nature of the trust network.Li and Kao [138]present a RS based on the trust of social networks;Through the trust computing,the quality and the veracity of peer produc-tion services can be appropriately assessed.The experimental re-sults show that the proposed RS can signiﬁcantly enhance the quality of peer production services.

Table3classiﬁes the current approaches to address user credi-bility and item reputation in social-based RS.

In the CFﬁeld,the trust of users is used to make predictions, weighting trust values.That is to say,the more trust a user has, the more important its ratings are for making predictions [58,112,239].In Ma et al.[145],they propose a probabilistic factor analysis framework,combining ratings and trusted friends;this framework can be applied to pure user-item rating matrix.

5.2.Content-basedﬁltering

Content-basedﬁltering(CBF)tries to recommend items to the active user similar to those rated positively in the past.It is based on the concept that items with similar attributes will be rated sim-ilarly[16,177,203].For example,if a user likes a web page with the words‘‘car’’,‘‘engine’’and‘‘gasoline’’,the CBF will recommend pages related to the automotive world.

CBF is becoming especially important as RS incorporate infor-mation on items from users working in web2.0environments, such as tags,posts,opinions and multimedia material.

Two challenging problems for content-basedﬁltering are lim-ited content analysis and overspecialization[3].Theﬁrst problem arises from the difﬁculty in extracting reliable automated informa-tion from various content(e.g.,images,video,audio and text), which can greatly reduce the quality of recommendations.The sec-ond problem(overspecialization)refers to the phenomenon in which users only receive recommendations for items that are very similar to items they liked or preferred;therefore,the users are not receiving recommendations for items that they might like but are unknown(e.g.,when a user only receives recommendations about ﬁctionﬁlms).Recommendations can be evaluated for novelty [32,105].

For CBF to operate,attributes of the items you wish to recom-mend must be extracted[176].Typically,a set of attributes is man-ually deﬁned for each item depending on its domain.In certain instances,such as when it is desired to recommend textual infor-mation,classic information retrieval techniques must be used to automatically deﬁne such attributes(e.g.,term frequency,inverse document frequency and normalization to page length).

Fig.8shows the CBF mechanism,which includes the following steps:(1)extract the attributes of items for recommendation,(2) compare the attributes of items with the preferences of the active

Table3

State of the art on trust and reputation.

User trust Item trust

Explicit

trust

systems The‘credibility’of users is calculated through explicit information of the rest of

users.[71,239,240].Services P2P usually implement this technique[138]

The‘reputation’of items is calculated by means of a feedback of users

who are asked about their opinions[114].E-commerce services often

use this technique

Implicit

trust

systems The‘credibility’of users is calculated through implicit information obtained in a

social network[59,150,200]

The‘reputation’of items is calculated studying how users work with

these items(for example,the number of times a song is played)

[58,122]

Memory

based

trust

The‘credibility’measure is calculated taking into account the users’ratings[112,127,145]

user,and(3)recommend items with characteristics thatﬁt the user’s interests.

When the attributes of the items and the user proﬁles are known,the key purpose for CBF[158]is to determine whether a user will like a speciﬁc item.This task is resolved traditionally by using heuristic methods[198,15,79]or classiﬁcation algorithms, such us:rule induction[65,119],nearest neighbors methods [236,27],Rocchio’s algorithm[131,16],linear classiﬁers[113],and probabilistic methods[175,160,84].

The pure CBF has several shortcomings[16,176,212]:

(a)In certain domains(e.g.,music,blogs,and videos),it is a

complicated task to generate the attributes for items.

(b)CBF suffers from an overspecialization problem because by

nature it tends to recommend the same types of items.

(c)It is more difﬁcult to acquire feedback from users because

with CBF,users do not typically rate the items(as in CF), and,therefore,it is not possible to determine whether the recommendation is correct.

Because of these shortcomings,it is rare toﬁnd a pure CBF implementation.It is more common to use the hybrid CBF/CF Burke2002.CF solves CBF’s problems because it can function in any domain;it is less affected by overspecialization;and it ac-quires feedback from users.CBF adds the following qualities to CF:improvement to the quality of the predictions,because they are calculated with more information,and reduced impact from the cold-start and sparsity problems.

CBF and CF can be combined in different ways[3].Fig.9shows the different alternatives.

Fig.9a shows the methods that calculate CBF and CF recom-mendations separately and subsequently combine them.Claypool et al.[]propose to use a weighted average for combining CBF and CF predictions depending on the type of prediction.In another study,Pazzani[177]proposes combining the CBF and CF recom-mendation lists by assigning the items scores according to their position on the lists.Additionally,Billsus and Pazzani[26]and Tran and Cohen[218]propose to select the CBF or CF prediction in accordance with the quality.

Fig.9b depicts the methods that incorporate CBF characteristics into the CF approach.Balabanovic and Shoham[16]maintain user proﬁles based on content analysis and directly compare the pro-ﬁles to determine similar users for CF recommendations.Good et al.[]construct specializedﬁlterbots using CBF techniques, which later act as neighbors in the CF stage.Melville et al.[157] propose to add predictions from the CBF into the ratting matrix employed by the CF.Li[136]modiﬁes the ratting matrix,which is input for the CF,by combining it with another matrix generated from clustering the items according to their attributes.In Hu and Pu[101],authors incorporate personality characteristics in the CF similarity measure to minimize the new-user problem.

Fig.9c illustrates the methods to construct a uniﬁed model with both CBF and CF characteristics.Basu et al.[19]propose using CBF and CF characteristics in a single rule-based classiﬁer. Popescul et al.[182]and Schein et al.[204]propose using prob-ability models to combine CBF and CF recommendations.In an-other studies[66,10,50],the authors employ Bayesian networks to combine CBF and CF characteristics and generate more accu-rate recommendations.Burke[45]and Middleton et al.[159] propose using knowledge-based techniques to solve the cold-start problem.

Fig.9d shows the methods that incorporate CF characteristics into a CBF approach.In Soboroff and Nicholas[211],the authors use LSI to create the user proﬁles used in CBF recommendations beginning with the CF ratting matrix.Mooney and Roy[160]use CF system predictions as input for CBF.

The current trend in CBF is to add social information to the items attributes,such as tags,comments,opinion,and social net-work sharing.Social tagging systems are the most popular because they allow users to annotate online resources with arbitrary labels, which produces rich information spaces(folksonomies).These new components have opened novel lines of RS research that can be di-vided into two categories:(1)tag recommendation systems and(2) use of tags in the recommendation process:

(1)RS tags attempt to provide personalized item recommenda-

tions to users through the most representative tags.In Jächke et al.[110],the authors compare different mecha-nisms for tags recommendations.Marinho and Schmidt-Thi-eme[147]improve tags recommendations by applying classic recommendation methods.Additionally,Landia and Anand[130]propose a method that combines clustering-based CBF with CF to suggest new tags to users.

(2)The methods using tags in the recommendation process

increase the capacity of traditional RS.Tso-Sutter et al.

[219]propose a generic method that allows tags to be incor-porated to standard CF algorithms.Bogers and Van Den Bosh

[39]examine how to incorporate the tags and other metada-

ta into a hybrid CBF/CF algorithm by replacing the tradi-tional user-based and item-based similarity measures by tag overlap.Gemmell et al.[83]propose a weighted hybrid recommender,wherein they combine the graph-based tag recommendations with user-based CF and item-based CF.

Gedikli and Jannach[81]propose to use tags as a means to express which features of an item users particularly like or dislike.In Gemmell et al.[82],the authors offer a hybrid RS,wherein they predict the user preferences for items by only consulting the user’s tagging history.

6.Additional recommender systems objectives

Commercial RS compete in the market by offering the best con-tent and quality in recommendations as well as greatest variety of services.Recommendations to user groups[108]facilitate joint recommendations to user groups(e.g.,a group of four friends who wish to choose a movie).For CF,four design approaches offer an opportunity for action:(1)acting into the similarity measures stage[168],(2)acquiring neighbors[37],(3)acquiring predictions [63],and(4)generating recommendations[17].Research results [168]indicate that the quality of the recommendations does not vary greatly between the different approaches,but the execution time is dramatically reduced as we advance when it is used(when the design of a similarity measure for groups is the most efﬁcient solution).

For the RS generated recommendations to be valuable for users, they must be explained well in a simple,compelling and accurate manner.The recommendation explanationﬁeld has been investi-gated with new developments in RS[91]until now[170].Tradi-tionally,the explanation type is divided into the following categories:(a)human style(user to user approach),(b)item style (item to item approach),(c)feature style(items features),and(d) hybrid.It also employs the use of conversational techniques[155] and incorporates geo-social information[235].

6.1.Recommending to groups of users

RS that consider groups of users[108]are starting to expand and to be used in different areas:tourism[14],music[55],TV [238],web[176].

Given the speciﬁc characteristics of the recommendation to groups,it is appropriate to establish a consensus for differentgroup semantics that formalize the agreements and disagreements among users[195].

With the aim of presenting the work carried out to date in a structured way,we provide a classiﬁcation of the recommendation to groups in CF RS.Fig.10graphically illustrates the four basic lev-els on which we can act in order to unify the group’s users’data with the objective of obtaining the data of the group of users:sim-ilarity metric,establishing the neighborhood,prediction phase, determination of recommended items.

In Fig.10,the individual members of a group are represented on the left,in grey;each graticule represents the matrix of ratings by the users(horizontal)on the items(vertical).The graph shows the four representative cases of tackling the solution to recommenda-tion by groups(one case for each matrix on the left of theﬁgure). The circles show key information:they indicate the CF process phase where the uniﬁcation is performed:‘‘n users?1group’’.

In theﬁrst case,at the top of the graph,the data uniﬁcation is performed in the prediction phase of the CF process:n individual predictions of n users of the group are combined in one prediction of the group(predictions aggregation).This approach has been used by Berkovsky and Freyne[22],García et al.[78]and Christen-sen and Schiafﬁno[63].

The second case acts on the sets of neighbors of the group’s users,by unifying them in one neighborhood for the whole group. This approach has been studied by Bobadilla et al.[37],proposing the intersection of a large number(k)of neighbors of each user of the group.

In the third case,the recommendations obtained for each indi-vidual user of the group are merged into one recommendation for the group.Baltrunas et al.[17]use rank aggregation of individual lists of recommendations.

The fourth case[168]uses a similarity metric that acts directly on the set of ratings of the group of users.This solution is the only one that directly provides a set of neighbors for the group of users.

A study exists[9]which,prior to any

poses,as a front-end,the incorporation

of missing information when dealing with

guistic preference relations.

6.2.Explaining recommendations

An important research subject in the

ing explanations that justify the recommendations

ceived.This is an important aspect of

maintaining a higher degree of user conﬁdence

erated by the system.

The type of explanations used thus

lows[170].

Human style explanations(user to user approach).For example, we recommend movie i because it was liked by the users who rated movies j,k,m,...very positively(j,k,m,...are movies rated well by the active user).

Item style explanations(item to item approach).For example,we recommend the vacation destination i because you liked the vacation destinations g,c,r,...(g,c,r,...are vacation destina-tions similar to i and rated well by the active user).

Feature style explanations(it is recommended based on items’features).For example,we recommend movie i because it was directed by director d,it features actors a,b,and it belongs to genre g(d,a,b,g are features the active user is interested in).

Hybrid methods.This category primarily includes the following: human/item,human/feature,feature/item,and human/feature/ item.

Additionally,in geo-social RS(Foursquare,Google latitude,etc.), location information exists that must be used in the recommenda-tion explanation mechanism[235].Geo-social RS typically adopt a hybrid human/item explanation method based on social,location and memory-based information.

A reference publication that is a helpful introduction to the RS explanations researchﬁeld has been published previously[91]. They explore the utility of explanations in CF RS,and they stated three key research questions:(1)What models and techniques are effective in supporting explanations?(2)Can explanation facil-ities increase the acceptance of CF RS?(3)Can explanation facilities increase theﬁltering performance of the CF RS users?To answer to theﬁrst question,they propose using rating histograms,indica-tions of past performance,comparisons to similar rated items, and use of domain speciﬁc content features.The results from the experiments conducted with RS users support an afﬁrmative re-sponse to the second question.The third question is unanswered

Fig.8.Content-basedﬁltering mechanism. Fig.9.Different alternatives for combining CF and CBF.because users performﬁltering based on many different channels of input.

A dynamic approach that favors the mechanisms for RS expla-nations includes using conversational techniques,such as the CCBR (conversational case-base reasoning),explained into McSherry [155].As CCBR they use an incremental nearest neighbor process based on the Pareto case dominance approach.In a different study [153],a dynamic approach is also adopted,but it employs a differ-ent perspective.Instead of attempting to justify a particular recom-mendation they focus on how explanations can help users to understand the recommendation opportunities that remain if the current recommendation should not meet their requirements. They generate compound critiques as explanations:Users have the opportunity to accept or critique recommendations.If they cri-tique a recommendation,the critique acts as aﬁlter over the remaining recommendations.

In a separate study[24],authors differentiate between the con-cepts promotion(increasing of the acceptance of the recommended item)and satisfaction(user satisfaction with the recommended item).They also produced better results by using the keyword style explanation(based on content data)compared with the neighbor style explanation(human style explanation).Authors propose a new classiﬁcation of the recommendation justiﬁcations:Keyword Style Explanation(for content-based RS),Neighbor Style Explana-key components:tag relevance,the degree to which a tag describes an item;and tag preference,the user’s sentiment toward a tag.

Fahri[73]provides a framework for organizing justiﬁcations, used to categorize explanations;they propose the categorization of the discourse:explicative,theoretical,pragmatic,ethical,moral, legal,aesthetic,and personal.Although this theoretical framework has not been used into the research literature,it can be used to de-sign new types of explanations.Hernando et al.[97]present a no-vel explanation technique based on the visualization of trees of items;these trees provide valuable information about the reliabil-ity of recommendations and the importance of the ratings the user has made.

The most relevant investigations that produce justiﬁcations in recommender systems include a study[187]wherein the authors design a new organization interface where results are grouped according to their tradeoff properties.They have developed a trust model for recommender agents based on the Pareto algorithm (excluding dominated categories).Symeonidis et al.[213]ﬁrst con-struct a feature proﬁle for the users to reveal their favorite features, later they group users into biclusters to exploit partial matching between de preferences of groups of users over groups of items. Additionally they propose a metric to measure the quality of justi-ﬁcations:the explain coverage ratio.In Symeonidis et al.[214]they use a prototype‘‘MoviExplain’’to put into the test the research

10.Classiﬁcation of the recommendations to groups in CF RS.Theﬁgure represents the four representative cases for approaching the solution to group recommendations.information,such as friends,followers,followed,both trusted and untrusted.Simultaneously,users aid in the collaborative inclusion of such information:blogs,tags,comments,photos and videos.(3) For the web3.0and the Internet of things,context-aware informa-tion from a variety of devices and sensors will be incorporated with the above information.Currently,geographic information is included,and the expected trend is gradual incorporation of information,such as radio frequency identiﬁcation(RFID)data,sur-veillance data,on-line health parameters and food and shopping habits,as well as teleoperation and telepresence.

Context-aware recommender systems[5,1],focus on additional contextual information,such as time,location,and wireless sensor networks[80].The contextual information can be obtained explic-itly,implicitly,using data mining or with a mixture of these meth-ods(hybrid).Currently,mobile applications increasingly use geographic information;this information enables geographic RS that can be considered as location-aware RS.For geographic RS [167,152],recommendations are typically generated by consider-ing the geographical position of the user that receives the recommendation.

This section provides an introduction of concepts,which are gaining popularity in the RS researchﬁeld:Internet of things,pri-vacy preservation,shilling attacks,new frameworks,etc.In this introduction,we provide a novel classiﬁcation for analyzing these RS concepts.Next,we will deal with the research on the loca-tion-aware RS,which may be regarded as theﬁrst steps for future RS based on Web3.0.Finally,we will describe the most signiﬁca-tive results on a promising researchﬁeld:the RS based on bio-in-spired models.

7.1.Introduction

There is a clear trend towards collection of implicit information instead of a traditional explicit evaluation of items by ratings. Last.Fm is a good example of this situation;the user ratings are in-ferred by the number of times they have heard each song.The same can be applied in a number of everyday situations,such as for access to web addresses,use of various public transport sys-tems,food purchased,access to sports facilities and access to learn-ing resources.

Incorporation of implicit information on the daily habits of users allows RS to use a variety of data;these data will be used in future CF processes,which are increasingly useful and accurate. Privacy and security considerations will be increasingly important with the widespread trend in using,with consent,devices and sen-sors for the Internet of things.

Privacy is an important issue for RS[23]because the systems contain information on large numbers of registered users.For pri-vacy preservation in RS,a certain level of uncertainty must be intro-duced into the predictions[156],primarily through tradeoffs between accuracy and privacy[146].Furthermore,privacy can be preserved when different RS companies share information(com-bining their data)[116,242].Privacy becomes more important as RS increasingly incorporate social information.

Because RS are often used in electronic commerce,unscrupulous producers mayﬁnd proﬁtable to shill RS by lying to the systems in order to have their products recommended more often than those of their competitors.RS can experience shilling attacks[128,57], which generate many positive ratings for a product,while products from competitors receive negative ratings.RS are still highly vul-nerable to such attacks[191].

Knowledge-basedﬁltering is emerging as an importantﬁeld of RS.Knowledge RS[46]‘‘use knowledge about users and products to pursue a knowledge-based approach to generating recommenda-tions,reasoning about what products meet the user’s requeri-ments’’.Recommendations are based on inferences about users needs and preferences.User models are based on knowledge struc-tures such as querys(preferred features por products)[109],cases (case-based reasoning)[44],constraints(constraint-based reason-ing)[74],ontologies[159],matching metrics and knowledge vec-tors[194],and social knowledge[53].

Workﬂow is a current knowledgeﬁeld where the user model is based on‘‘users-roles-tasks reference information that describes which member plays which roles or fulﬁlls which tasks’’[245,246].Peer-to-peer(P2P)networks are other current knowl-edgeﬁeld,where user information is based on the distributed information existing from each peer and the set of peers who may need her[247].

Gradual incorporation of different types of information(e.g.,ex-plicit ratings,social relations,user contents,locations,use trends, knowledge-based information)has forced RS to use hybrid ap-proaches.Once the memory-based,social and location-aware methods and algorithms are consolidated,the evolution of RS dem-onstrates a clear trend toward combining existing collaborative methods.

The latest research in the CFﬁeld has generated only modest improvements for predictions and recommendations from a single type of information(e.g.,when the only information used is user ratings,information from social relations,or item content).The re-sults improve further when several algorithms are combined with their respective data types.A growing number of publications ad-dress hybrid approaches that use current databases to simulta-neously incorporate memory-based,social and content-based information.

To unify the above concepts,Fig.11provides an original taxon-omy for RS.The taxonomy is classiﬁed depending on the nature of the data rather than according to the methods and algorithms used.The core of the taxonomy focuses on data classiﬁcation by three factors:(1)the target of the data:user or item;(2)mode of acquisition:explicit(i.e.,ratings to items made by users)or impli-cit(e.g.,number of times a user has heard a song);and(3)informa-tion level:memory,content or social context.

Fig.11shows the recommender methods and algorithms(la-beled as‘‘collaborativeﬁltering algorithms’’).Depending on the information type in each RS database,it adopts a hybridﬁltering approach.Each hybrid approach will use an appropriate subset of algorithms to consider processing of existing information in a coor-dinated manner.Future developments will include different rec-ommendation frameworks that address the most common situations.These frameworks allow RS to incorporate the CF kernel with the most appropriate recommendations methods based on the available information in a simple and straightforward manner.

At higher levels(prediction and recommendation),Fig.11 incorporates current evaluation quality measures,such as those for diversity and novelty.The importance of such measures,and measures developed in the future will grow as users demand novel, stable and less predictable recommendations.

7.2.Location-aware recommender systems

Due to the increasing use of mobile devices,location-aware sys-tems are becoming more widespread.These systems show a ten-dency towards their consolidation as web3.0services and this naturally leads to location-aware CF and location-aware RS,which can be called geographic CF and geographic RS.

We introduce a classiﬁcation for geographic CF RS and focus on the most relevant section of the classiﬁcation obtained.Table4 establishes the different possibilities of tackling a geographic RS according to the nature of the ratings made(‘‘rating stage’’)and the recommendation process followed(‘‘recommendation stage’’).‘‘User’’indicates that the rating and/or recommendation are made without having or using the user’s Geographic Information(GI).Similarly,‘‘Item’’indicates that the rating and/or recommendation are made without having or using the item’s GI.In the cases la-beled as‘‘User g’’and‘‘Item g’’the GI is used.

The cases identiﬁed are:

RS:Traditional RS,in which ratings and recommendations are made without using geographical information.

RS+G:Traditional RS,which also contributes the item’s geo-graphical position.These RS cannot be regarded as geographic RS,as the GI does not play a part in the recommendation process.

GRS:This group of Geographic RS is most likely to become pop-ular in the near future.In these,ratings are made in a traditional way,whilst recommendations are made by considering the geo-graphical position of the user to whom the recommendation is to be made.A representative example is that of a RS for restau-rants;the users rate a restaurant using very diverse concepts, which do not include the distance at the time of voting between the user and the restaurant.However,users of a Geographic RS expects a restaurant to be recommended to them not only because of good ratings from similar users(k-neighbors),but also according to the distance between their current position and that of the restaurant.Other possible examples are RS for cinemas,pubs,supermarkets,cultural activities in a city,lan-guage learning centers,gyms and sports clubs,etc.

GRS+:In this case,users establish ratings on items by weighting the distance between them and the items rated.In this type of geographic RS two possibilities can be established:

1.Hybrid CF/Demographicﬁltering:Each item accepts a max-

imum of one vote per user,to which the geographical posi-

tion from which it has been issued is associated.

2.Geographic RS where each item accepts more than one rat-

ing for each user,depending on the geographical position

from which each rating is made.

3.The hybrid RS in case1respond to regional or national geo-

graphical approaches,in which recommendations can be

established according to weighting between the similarity

of the votes(CF)and their origin.This type of GRS may

be regarded as an extended case of hybrid CF/demographic

ﬁltering,in which the GI is given for each vote instead of for

each user.

From a theoretical point of view,Type2GRS+are the most com-plete;however,from a practical point of view,they involve a semantic difﬁculty in the item rating process,which makes their use very difﬁcult.Rating items in this GRS+involves that each user can rate items according to the relative distances between the user and the items.In this way,a user can rate a restaurant from their home differently to how they would rate it from their workplace; and when the distances are very different,the ratings are also likely to be so.The mental process would be something like this: I am1km from the restaurant and I rate very positively travelling 1km to go to that restaurant which I think is good;but after some time,the same user,who is at work,24km away from the restau-rant,could cast a vote indicating they do not consider it to be po-sitive to travel24km to go to the restaurant even if they think it is good.

In summary,GRS+have the advantage that they accept a wider variety of ratings and that these also contain the relative impor-tance that each user gives to the items according to the distance re-quired to access them.The disadvantage is that it is difﬁcult to involve users in a particularly complex and demanding ratings process.

This subsection focuses on the GRS-type geographic CF RS.At present,there are few publications regarding GI-based RS;This is due,to a great extent,to the lack of public databases that include ratings and geographic positions capable of being combined in an RS.Some of the publications that focus more closely on theﬁeld are as follows:

Martinez et al.[149]and Biuk-Aghai et al.[28]are examples of the RS+G group.In Schlieder[205],they propose a novel approach for modeling the collaborative semantics of geographic folksono-mies.This approach is based on multi-object tagging,that is,the analysis of tags that users assign to composite objects.This paper is based on the concept of groups of people who share a common geospatial feature data dictionary(including deﬁnitions of feature relationships)and a common metadata schema.

Wan-Shiou et al.[225]can be considered as a hybrid content based/geographic RS.The core of the system is a hybrid content based/geographic recommendation mechanism that analyzes a customer’s history and position so that vendor information can be ranked according to the match with the preferences of a customer.

Matyas and Schlieder[152]show a collaborative system that we could situate between a RS and a GRS.In this case,the users’ratings are taken based on the photos they have downloaded from a Web2.0and the photos they have uploaded to the same Web (the photos have a GPS address associated to them).After this,a search of k-neighborhoods based on this data is carried out.The recommendation process does not take into account the user’s position.

It is possible to collect travel GPS traces from users and use the database to generate recommendations[249].The travel GPS traces can be reinforced with social information based on friends [250].Both papers can be classiﬁed as GRS+.

7.3.Bio-inspired approaches

Much of the proposed model-based RS are based on bio-in-spired approaches,which primarily use Genetic Algorithms(GAs) and Neural Networks(NNs).Models have also been proposed based on Artiﬁcial Immune Networks(AINs).

GA are heuristic approaches based on evolutionary principles such as natural selection and survival of theﬁtest.GA have mainly been used in two aspects of RS:clustering[120,243]and hybrid user models[76,99,7].A common technique to improve the fea-tures of RS consists of initially carrying out a clustering on all of the users,in such a way that a group of classes of similar users is obtained,after this,the desired CF techniques can be applied to each of the clusters,obtaining similar results but in much shorter calculation times;It is usual to use common genetic clustering algorithms such as GA-based K-means[121].

The RS hybrid user models commonly use a combination of CF with demographicﬁltering or CF with content basedﬁltering,to exploit merits of each one of these techniques.In these cases,the chromosome structure can easily contain the demographic charac-teristics and/or those related to content-basedﬁltering.

In order to tackle location-based advertisement,Dao et al.[68] propose a model-based CF using GA.They combine both user’s preferences and interaction context.Bobadilla et al.[33]use GA to create a similarity metric,weighting a set of very simple similar-ity measures.Hwang et al.[106]employ a GA to learn personal preferences of customers.

NN is a model based on the observed behavior of biological neu-rons.This model,intended to simulate the way the brain processes information,enables the computer to‘‘learn’’to a certain degree.A NN typically consists of a number of interconnected nodes.Each handles a designated sphere of knowledge,and has several inputs from the network.Based on the inputs it gets,a node can‘‘learn’’about the relationships between sets of data,pattern,and,based upon operational feedback,are molded into the pattern required to generate the required results.The RS most relevant research available in which NN usually fo-cuses is hybrid RS,in which NN are used for learn users proﬁles; NN have also been used in the clustering processes of some RS.

The hybrid approaches enable NN to act on the additional infor-mation to the ratings.In Ren et al.[192]they propose a hybrid rec-ommender approach that employs Widrow-Hoff[229]algorithm to learn each user’s proﬁle from the contents of rated items.This improves the granularity of the user proﬁling.In Christakou and Stafylopatis[62]they use a combination of content-based and CF in order to construct a system that provides more precise recom-mendations concerning movies.In Lee and Woo[133]ﬁrst,all users are segmented by demographic characteristics and users

in Fig.11.Recommender systems taxonomy.

each segment are clustered according to the preference of items using the Self-Organizing Map(SOM)NN.Kohonon’s SOMs are a type of unsupervised learning;their goal is to discover some underlying structure of the data.

Two alternative NN uses are presented in Huang et al.[103]and Roh et al.[193].In theﬁrst case paper,authors use a training back-propagation NN for generating association rules that are mined from a transactional database;in the second paper,authors pro-pose a model that combines a CF algorithm with two machine learning processes:SOM and Case Based Reasoning(CBR)by changing an unsupervised clustering problem into a supervised user preference reasoning problem.

Neuro-fuzzy inference has been used in Sevarac et al.[207]to create pedagogical rules in e-learning.A new cold-start similarity measure has been perfected in Bobadilla et al.[36]using optimiza-tion based on neural learning.

Artiﬁcial immune systems are distributed and adaptive systems using the models and principles derived form the human immune system.They model the defence system which can protect our body against infections.In order to tackle the RS sparsity problem and to make algorithms more scalable,Acilar and Arslan[2]pres-ent a new CF model based on the AIN Algorithm(aiNet).AIN were previously proposed to general recommendations[49]and to rec-ommend web sites[161].

8.Related works and original contributions of the paper

As CF has become more complex,different survey papers have been published in this area.Schafer et al.[203]introduces the core concepts of CF:the theory and practice,the rating systems and their acquisition,evaluation,interaction interfaces and privacy is-sues.Candillier et al.[51]review the main CFﬁltering methods and compare their results.

Su and Khoshgoftaar[212]presents a survey of CF techniques. Authors introduce the theory on CF and concisely deal with the main challenges:sparsity,scalability,synonymy,gray sheep,shil-ling attacks,privacy,etc.They also expose an overview table of CF techniques.

Park et al.[171]review210papers on RS and classiﬁes them by the year and journal of the publication,their applicationﬁelds,and their data mining techniques.Additionaly,they categorized the pa-pers into eight applicationﬁelds(ﬁlms,music,etc.).

A review in RS algorithms is presented in[141].This paper fo-cuses on explaining carefully how the most used algorithms in RS work.The paper presents also the basic concepts of CF and their evaluation metrics,dimensionality reduction techniques,diffu-sion-based methods,socialﬁltering and meta approaches.

Our survey tries to include the most novel issues that have not been dealt carefully in the previous papers.Next,we will stand out the most outstanding features of this survey:

Uses a methodology for selecting the most suitable papers in the RS,standing out the latest and most cited papers in the area of RS.

Provides an updated overview table of the most used RS public databases,including tags and friend relations information.

Studies the cold-start problem inherent to all the RS.

Presents a novel overview table informing both the classical similarity measure and those which have recently been pro-posed.It includes both the tailored metrics for cold-start users and the general-purpose metrics.Besides,we show the quality measures obtained when evaluating such metrics.

Includes the recent quality measurements,beyond accuracy,to evaluate RS:novelty,diversity and stability.Additionaly,we include a reliability measure associated to predictions and recommendations.

Provides a comprehensive survey on socialﬁltering,presenting

a novel overview table on trust,reputation and credibility.

Introduces the content-basedﬁltering from a modern perspec-tive standing out its application for dealing with social informa-tion,such as social tagging.

Presents a summary of the most relevant contributions in the RS for group of users.We will show a novel classiﬁcation for the existing methods.

Deals with a fast growing RSﬁeld:the location-aware RS,based on geographic information.This section is estructured with the help of a novel geographic RS classiﬁcation table.

Summarizes the most relevant contributions on the use of bio-inspired approaches.

Describes the RS trends to implicitally collect data(specially those derived from the use of Internet of things).

Provides an RS taxonomy for classifying the RS through three factors:source of data(traditional web,social web2.0,Internet of things/web3.0);target of data(users,items);method for extracting data(explicit,implicit).

9.Conclusions

Recommender systems are proving to be a useful tool for addressing a portion of the information overload phenomenon from the Internet.Its evolution has accompanied the evolution of the web.Theﬁrst generation of recommender systems used tradi-tional websites to collect information from the following three sources:(a)content-based data from purchased or used products, (b)demographic data collected in users’records,and(c)mem-ory-based data collected from users’item preferences.The second generation of recommender systems,extensively use the web2.0 by gathering social information(e.g.,friends,followers,followed, trusted users,untrusted users).The third generation of recom-mender systems will use the web3.0through information pro-vided by the integrated devices on the Internet.The use of location information already incorporated in many recommender systems will be followed by data from devices and sensors,which will be widely used(e.g.,real-time health signals,RFID,food habits, online local weather parameters such as temperature and pressure).

Theﬁrsts recommender systems were focused on improving recommendation accuracy throughﬁltering.Most memory-based methods and algorithms were developed and optimized in this context(e.g.,k NN metrics,aggregation approaches,singular value decomposition,diffusion-based methods,etc.).At this stage,hybrid approaches(primarily collaborative–demographic and collabora-tive–contentﬁltering)improved the quality of the recommenda-tions.In the second stage,algorithms that included social information with previous hybrid approaches were adapted and developed(e.g.,trust-aware algorithms,social adaptive ap-proaches,social networks analysis,etc.).Currently,the hybrid ensemble algorithms incorporate location information into exist-ing recommendation algorithms.

Evaluation of the predictions and recommendations has evolved since the origins of recommender systems,which weighted prediction errors(accuracy)heavily.They also recognized the

Table4

Geographic collaborativeﬁltering recommender systems classiﬁcation.

Rating stage Recommendation stage User GI

Item Item g Item Item g

User RS/GRS–RS RS+G Not User g–GRS+–GRS/GRS+Yes Item GI Not Yes Not Yesconvenience of evaluating the quality of the top n recommenda-tions as a set;evaluation of the top n recommendations as a ranked list was then incorporated.Currently,there is a tendency to assess new evaluation measures,such as diversity and novelty.

Future research will concentrate on advancing the existing methods and algorithms to improve the quality of recommender systems predictions and recommendations.Simultaneously,new lines of research will be developed forﬁelds and aims,such as on:(1)proper combination of existing recommendation methods that use different types of available information,(2)to get the maximum use of the individual potential of various sensors and devices on the Internet of things,(3)acquisition and integration of trends related to the habits,consumption and tastes of individ-ual users in the recommendation process,(4)data mining from RS databases for non-recommendation uses(e.g.,market research, general trends,visualization of differential characteristics of demo-graphic groups),(5)enabling security and privacy for recom-mender systems processes,(6)new evaluation measures and developing a standard for non-standardized evaluation measures, and(7)designingﬂexible frameworks for automated analysis of heterogeneous data.

References

[1]S.Abbar,M.Bouzeghoub,S.Lopez,Context-aware recommender systems:a

service oriented approach,in:Proceedings of the3rd International Workshop on Personalized Access,Proﬁle Management and Context Awareness in Databases,2009.

[2]A.M.Acilar,A.Arslan,A collaborativeﬁltering method based on artiﬁcial

immune network,Expert Systems with Applications36(4)(2008)8324–8332.

[3]G.Adomavicius,A.Tuzhilin,Toward the next generation of recommender

systems:a survey of the state-of-the-art and possible extensions,IEEE Transactions on Knowledge and Data Engineering17(6)(2005)734–749. [4]G.Adomavicius,J.Zhang,On the stability of recommendations algorithms,in:

ACM Conference on Recommender Systems,2010,pp.47–54.

[5]G.Adomavicius,A.Tuzhilin,Context-Aware recommender Systems,in:F.

Ricci,et al.(Ed.),Recommender Systems Handbook,2011,pp.217–253. [6]H.J.Ahn,A new similarity measure for collaborativeﬁltering to alleviate the

new user cold-starting problem,Information Sciences178(2008)37–51. [7]M.Y.H.Al-Shamri,K.K.Bharadwaj,Fuzzy-genetic approach to recommender

systems based on a novel hybrid user model,Expert Systems with Applications35(3)(2008)1386–1399.

[8]J.Al-Sharawneh,M.A.Williams,Credibility-aware Web-based social network

recommender:follow the leader,in:Proceedings of the2010ACM Conference on Recommender Systems,2010,pp.1–8.

[9]S.Alonso,F.J.Cabrerizo,F.Chiclana,F.Herrera,E.Herrera-Viedma,Group

decision making with incomplete fuzzy linguistic preference relations, International Journal of Intelligent Systems24(2009)201–222.

[10]A.Ansari,S.Essegaier,R.Kohli,Internet recommendation systems,Journal of

Marketing Research37(3)(2000)363–375.

[11]N.Antonopoulus,J.Salter,Cinema screen recommender agent:combining

collaborative and content-basedﬁltering,IEEE Intelligent Systems(2006)35–

41.

[12]P.Antunes,V.Herskovic,S.F.Ochoa,J.A.Pino,Structuring dimensions for

collaborative systems evaluation,ACM Computing Surveys44(2)(2012).

Article8.

[13]O.Arazy,N.Kumar,B.Shapira,Improving Social Recommender Systems,

Journal IT Professional11(4)(2009)31–37.

[14]L.Ardissono, A.Goy,G.Petrone,M.Segnan,P.Torasso,INTRIGUE:

Personalized recommendation of tourist attractions for desktop and handset devices,Applied Artiﬁcial Intelligence17(8-9)(2003)687–714. [15]R.Baeza-Yates, B.Ribeiro-Neto,Modern Information Retrieval,Addison-

Wesley,1999.

[16]M.Balabanovic,Y.Shoham,Content-based,collaborative recommendation,

Communications of the ACM40(3)(1997)66–72.

[17]L.Baltrunas,T.Makcinskas, F.Ricci,Group recommendation with rank

aggregation and collaborativeﬁltering,in:Proceedings of the2010ACM Conference on Recommender Systems,2010,pp.119–126.

[18]A.B.Barragáns-Martı´nez,E.Costa-Montenegro,J.C.Burguillo,M.Rey-López,

F.A.Mikic-Fonte, A.Peleteiro,A hybrid content-based and item-based

collaborativeﬁltering approach to recommend TV programs enhanced with singular value decomposition,Information Sciences180(22)(2010)4290–4311.

[19]C.Basu,H.Hirsh,W.Cohen,Recommendation as classiﬁcation:using social

and content-based information in recommendation,in:Proceedings of the Fifteenth National Conference on Artiﬁcial Intelligence,1998,pp.714–720.[20]P.Bedi,R.Sharma,Trust based recommender system using ant colony for

trust computation,Expert Systems with Applications39(1)(2012)1183–1190.

[21]Y.Bengio,Y.Grandvalet,No umbiased estimator of the variance of k-fold

cross-validation,Journal of Machine Learning Research5(2004)10–1105.

[22]S.Berkovsky,J.Freyne,Group-based recipe recommendations:analysis of

data aggregation strategies,in:Proceedings of the2010ACM Conference on Recommender Systems,2010,pp.111–118.

[23]A.Bilge,H.Polat,An improved privacy-preserving DWT-based collaborative

ﬁltering scheme,Experts Systems with Applications39(3)(2012)3654–3841.

[24]M.Bilgic,R.Mooney,Explanation for recommender systems:satisfaction vs.

promotion,in:Next Stage of Recommender Systems Research Workshop(IUI conference),2005,pp.13–18.

[25]D.Billsus,M.Pazzani,A personal news agent that talks,learns and explains,

in:Proc.Auton.Agents Conf.,1999,pp.268–275.

[26]D.Billsus,M.Pazzani,User modeling for adaptive news access,User Modeling

and User-Adapted Interaction10(2–3)(2000)147–180.

[27]D.Billsus,M.Pazzani,J.Chen,A learning agent for wireless news access,in:

Proceedings of the International Conference on Intelligent User Interfaces, 2002,pp.33–36.

[28]R.P.Biuk-Aghai,S.Fong,S.Yain-Whar,Design of a recommender system for

mobile tourism multimedia selection,in:2nd International Conference on Internet Multimedia Services Architecture and Applications(IMSAA),2008, pp.1–6.

[29]J.Bobadilla, F.Serradilla,The effect of sparsity on collaborativeﬁltering

metrics,in:Australian Database Conference,2009,pp.9–17.

[30]J.Bobadilla, F.Serradilla, A.Hernando,Collaborativeﬁltering adapted to

recommender systems of e-learning,Knowledge Based Systems22(2009) 261–265.

[31]J.Bobadilla,F.Serradilla,J.Bernal,A new collaborativeﬁltering metric that

improves the behavior of recommender systems,Knowledge Based Systems 23(2010)520–528.

[32]J.Bobadilla,A.Hernando,F.Ortega,J.Bernal,A framework for collaborative

ﬁltering recommender systems,Expert Systems with Applications38(12) (2011)14609–14623.

[33]J.Bobadilla,F.Ortega,A.Hernando,J.Alcalá,Improving collaborativeﬁltering

recommender systems results and performance using genetic algorithms, Knowledge Based Systems24(8)(2011)1310–1316.

[34]J.Bobadilla,A.Hernando,F.Ortega,A.Gutiérrez,Collaborativeﬁltering based

on signiﬁcances,Information Sciences185(1)(2012)1–17.

[35]J.Bobadilla, F.Ortega, A.Hernando,A collaborativeﬁltering similarity

measure based on singularities,Information Processing and Management 48(2)(2012)204–217.

[36]J.Bobadilla, F.Ortega, A.Hernando,J.Bernal,A collaborativeﬁltering

approach to mitigate the new user cold start problem,Knowledge Based Systems26(2012)225–238.

[37]J.Bobadilla, F.Ortega, A.Hernando,J.Bernal,Generalization of

recommender systems:collaborativeﬁltering extended to groups of users and restricted to groups of items,Expert Systems with Applications39(2012) 172–186.

[38]J.Bobadilla,F.Ortega,A.Hernando,A.Arroyo,A balanced memory-based

collaborativeﬁltering similarity measure,International Journal of Intelligent Systems27(10)(2013)939–946.

[39]T.Bogers,A.Van Den Bosch,Collaborative and content-basedﬁltering for

item recommendation on social bookmarking websites,in:Proceedings of the 2009ACM Conference on Recommender Systems,2009,pp.9–16.

[40]P.Bonhard,Who do trust?Combining recommender systems and social

networking for better advice,in:International Conference on Intelligent User Interfaces,2005.

[41]P.Bonhard,M.A.Sasse,‘Knowing me,knowing you’—Using proﬁles and social

networking to improve recommender Systems,BT Technology Journal24(3) (2006)84–98.

[42]P.Borzymek,M.Sydow,A.Wierbicki,Enriching trust prediction model in

social network with user rating similarity,in:Proceedings of the2009 International Conference on Computational Aspects of Social Network,2009, pp.40–47.

[43]J.S.Breese, D.Heckerman, C.Kadie,Empirical analysis of predictive

algorithms for collaborativeﬁltering,in:14th Conference on Uncertainty in Artiﬁcial Intelligence,1998,pp.43–52.

[44]D.Bridge,M.H.Goker,L.McGinty, B.Smyth,Case-based recommender

systems,The Knowledge Engineering Review20(3)(2005)315–320. [45]R.Burke,Encyclopedia of library and information systems,in:A.Kent(Ed.),

vol.69(Suppl.32),Marcel Dekker,2000(Chapter:Knowledge-Based Recommender Systems).

[46]R.Burke,A case-based reasoning approach to collaborativeﬁltering,in:

EWCBR2000,2000,pp.370–379.

[47]R.Burke,Hybrid recommender systems:survey and experiments,User

Modeling and User-Adapted Interaction12(4)(2002)331–370.

[48]F.Cacheda,V.Carneiro, D.Fernández,V.Formoso,Comparison of

collaborativeﬁltering algorithms:limitations of current techniques and proposals for scalable,high-performance recommender Systems,ACM Transactions on the Web5(1)(2011).Article2.

[49]S.Caizer,U.Aickelin,A recommender system based on idiotypic artiﬁcial

immune networks,Journal of Mathematics,Models and Algorithms4(2) (2005)181–198.[50]L.M.Campos,J.M.Fernández-Luna,J.F.Huete,M.A.Rueda-Morales,

Combining content-based and collaborative recommendations:a hybrid approach based on Bayesian Networks,International Journal of Approximate Reasoning51(7)(2010)785–799.

[51]L.Candillier,F.Meyer,M.Boullé,Comparing state-of-the-art collaborative

ﬁltering systems,Lecture Notes in Computer Sciece4571(2007)548–562.

[52]F.Carmagnola, F.Vernero,P.Grillo,SoNARS:a social networks-based

algorithm for social recommender systems,in:Proceedings of the17th International Conference on User Modeling,Adaptation,and Personalization: Formerly UM and AH,2009,pp.223–234.

[53]W.Carrer-Neto,M.L.Hernández-Alcaraz,R.Valencia-Garcı´a, F.Garcı´a-

Sánchez,Social knowledge-based recommender system,Application to the movies domain.Expert Systems with Applications39(12)(2012)10990–11000.

[54]J.J.Castro-Sanchez,R.Miguel,D.Vallejo,L.M.López-López,A highly adaptive

recommender system based on fuzzy logic for B2C e-commerce portals, Expert Systems with Applications38(3)(2011)2441–2454.

[55]D.Chao,J.Balthrop,S.Forrest,Adaptive radio:achieving consensus using

negative preferences,in:International ACM SIGGROUP Conference on Supporting Group Work,2005,pp.120–123.

[56]T.Chen,L.He,Collaborativeﬁltering based on demographic attribute vector,

in:Proceedings of the International Conference on Future Computer and Communication,2009,pp.225–229.

[57]P.A.Chirita,W.Nejdl, C.Zamﬁr,Preventing shilling attacks in online

recommender systems,in:Workshop on Web Information and Data Management,2005,pp.67–74.

[58]J.Cho,K.Kwon,Y.Park,Q-rater:a collaborative reputation system based on

source credibility theory,Expert Systems with Applications36(2009)3751–3760.

[59]S.B.Cho,J.H.Hong,M.H.Park,Location-based recommendation system using

Bayesian user’s preference model in mobile devices,Lecture Notes in Computer Science4611(2007)1130–1139.

[60]K.Choi,D.Yoo,G.Kim,Y.Suh,A hybrid online-product recommendation

system:combining implicit rating-based collaborativeﬁltering and sequential pattern analysis.Electronic Commerce Research and Applications,in press,doi:10.1016/j.elerap.2012.02.004.

[61]K.Choi,Y.Suh,A new similarity fuction for selecting neighbors for each target

item in collaborativeﬁltering,Knowledge Based Systems37(2013)146–153.

[62]C.Christakou,A.Stafylopatis,A hybrid movie recommender system based on

neural networks,in:International Conference on Intelligent Systems Design and Applications,2005,pp.500–505.

[63]I.A.Christensen,S.Schiafﬁno,Entertainment recommender systems for group

of users,Expert Systems with Applications38(2011)14127–14135.

[]M.Claypool, A.Gokhale,T.Miranda,P.Murnikov, D.Netes,M.Sartin,

Combining content-based and collaborativeﬁlters in an online newspaper, in:Proceedings of ACM SIGIR Workshop on Recommender Systems,1999,pp.

40–48.

[65]W.Cohen,Fast effective rule induction,in:Proceedings of the Twelfth

International Conference on Machine Learning,1995,pp.115–123.

[66]M.Condliff,D.Lewis,D.Madigan,C.Posse,Bayesian mixed-effects models for

recommender systems,in:ACM SIGIR’99Workshop on Recommender Systems:Algorithms and Evaluation,1999,pp.23–30.

[67]E.Costa-Montenegro,A.B.Barragáns-Martı´nez,M.Rey-López,Which App?A

recommender system of applications in markets:implementation of the service for monitoring users’interaction,Expert Systems with Applications 39(10)(2012)9367–9375.

[68]T.H.Dao,S.R.Jeong,H.Ahn,A novel recommendation model of location-based

advertising:context-aware collaborativeﬁltering using GA approach,Expert Systems with Applications39(3)(2012)3731–3739.

[69]M.Dell’amico,L.Capra,SOFIA:socialﬁltering for robust recommendation,

IFIP Advances in Information and Communication Technology263(2008) 135–150.

[70]T.Dubois,J.Golbeck,J.Kleint,A.Srinivasan,Improving recommendation

accuracy by clustering social networks with trust,in:Proceedings of the2009 ACM Conference on Recommender Systems,2009,pp.1–8.

[71]M.Ekström,H.Björnsson,C.Nass,A reputation mechanism for business-to-

business electronic commerce that accounts for rater credibility,Journal of Organizational Computing and Electronic Commerce15(1)(2005)1–18. [72]I.Esslimani,A.Brun,A.Boyer,From social networks to behavioral networks in

recommender systems,in:Proceedings of the2009International Conference on Advances in Social Network Analysis and Mining,2009,pp.143–148. [73]Y.Fahri,A Framework for Organizing Justiﬁcations for Strategic use in

Adaptive Iteraction Contexts,ECIS,2008.Article250.

[74]A.Felfernig,R.Burke,Constraint-based recommender systems:technologies

and research issues,in:10th International Conference on Electronic Commerce,2008(Article No.3).

[75]L.Fengkun,J.L.Hong,Use of social network information to enhance

collaborativeﬁltering performance,Expert Systems with Applications37

(7)(2010)4772–4778.

[76]L.Q.Gao,C.Li,Hybrid personalizad recommended model based on genetic

algorithm,in:International Conference on Wireless Communication, Networks and Mobile Computing,2008,pp.9215–9218.

[77]M.Gao,Z.Wu, F.Jiang,Userrank for item-based collaborativeﬁltering

recommendation,Information Processing Letters111(9)(2011)440–446.

[78]I.Garcı´a,L.Sebastia,E.Onaindia,On the design of individual and group

recommender systems for tourism,Expert Systems with Applications38 (2011)7683–7692.

[79]R.Garcı´a,X.Amatriain,Weighted content based methods for recommending

connections in online social networks,in:Proceedings of the2010ACM conference on Recommender Systems,2010,pp.68–71.

[80]D.Gavalas,M.Kenteris,A web-based pervasive recommendation system for

mobile tourist guides,Personal and Ubiquitous Computing15(7)(2011) 759–770.

[81]F.Gedikli,D.Jannach,Rating items by rating tags,in:Proceedings of the2010

ACM Conference on Recommender Systems,2010,pp.25–32.

[82]J.Gemmell,T.Schimoler,B.Mobasher,R.Burke,Resource recommendation

for social tagging:a multi-channel hybrid approach,in:Proceedings of the 2010ACM Conference on Recommender Systems,2010,pp.60–67.

[83]J.Gemmell,T.Schimoler,M.Ramezani,L.Christiansen, B.Mobasher,

Improving FolkRank with item-based collaborativeﬁltering,in:Proceedings of the2009ACM conference on Recommender Systems,2009,pp.17–24. [84]M.Gemmis,P.Lops,G.Semeraro,P.Basile,Integrating tags in a semantic

content-based recommender,in:Proceedings of the2008ACM conference on Recommender Systems,2008,pp.163–170.

[85]T.George,S.Meregu,A scalable collaborativeﬁltering Framework base don

co-clustering,in:IEEE International Conference on Data Mining(ICDM),2005, pp.625–628.

[86]J.Golbeck,U.Kuter,The ripple effect:change in trust and its impact over a

social network,in:Computing with Social Trust,Human–Computer Interaction Series,Part II,2009,pp.169–181(Chapter7).

[87]K.Goldberg,T.Roeder, D.Gupta,C.Perkins,Eigentaste:a constant time

collaborativeﬁltering algorithm,Information Retrieval4(2)(2001)133–151.

[88]R.González-Crespo,O.Sanjuán-Martı´nez,J.Manuel-Cueva,B.Cristina-Pelayo,

J.E.Labra-Gayo,P.Ordoñez,Recommendation system based on user interaction data applied to intelligent electronic books,Computers in Human Behavior27(4)(2011)1445–1449.

[]N.Good,J.B.Schafer,J.A.Konstan,A.Borchers,B.Sarwar,J.L.Herlocker,J.

Riedl,in:Proceedings of the Sixteenth National Conference on Artiﬁcial Intelligence and the Eleventh Innovative Applications of Artiﬁcial Intelligence Conference Innovative Applications of Artiﬁcial Intelligence,1999,pp.439–446.

[90]A.Gunawardana,G.Shani,A survey of accuracy evaluation metrics of

recommender tasks,Journal of Machine Learning Reearch10(2009)2935–2962.

[91]J.L.Herlocker,J.A.Konstan,J.Riedl,Explaining collaborativeﬁltering

recommendations,in:ACM Conference on Computer Supported Cooperative Work(CSCW),2000,pp.241–250.

[92]J.L.Herlocker,J.A.Konstan,A.L.Borchers,J.T.Riedl,An algorithmic framework

for performing collaborativeﬁltering,in:Proceedings of the22nd Annual International ACM SIGIR Conference on Research and Development in Information Retrieval,1999,pp.230–237.

[93]J.L.Herlocker,J.A.Konstan,J.T.Riedl,An empirical analysis of design choices

in neighborhood-based collaborativeﬁltering algorithms,Information Retrieval5(2002)287–310.

[94]J.L.Herlocker,J.A.Konstan,J.T.Riedl,L.G.Terveen,Evaluating collaborative

ﬁltering recommender systems,ACM Transactions on Information Systems 22(1)(2004)5–53.

[95]F.Hernández, E.Gaudioso,Evaluation of recommender systems:a new

approach,Expert Systems with Applications35(2008)790–804.

[96]A.Hernando,J.Bobadilla, F.Ortega,J.Tejedor,Incorporating reliability

measurements into the predictions of a recommender systems.Information Sciences,in press,doi:10.1016/j.ins.2013.03.018.

[97]A.Hernando,J.Bobadilla, F.Ortega, A.Gutiérrez,Trees for explaining

recommendations made through collaborativeﬁltering,Information Sciences218(2013)1–16.

[98]K.Heung-Nam,E.S.Abdulmotaleb,J.Geun-Sik,Collaborative error-reﬂected

models for cold-start recommender systems,Decision Support Systems51(3) (2011)519–531.

[99]Y.Ho,S.Fong,Z.Yan,A hybrid ga-based collaborativeﬁltering model for

online recommenders,in:International Conference on e-Business,2007,pp.

200–203.

[100]L.Hossain,D.Fazio,The social networks of collaborative process,The Journal of High Technology Management Research20(2)(2009)119–130.

[101]H.R.Hu,P.Pu,Using personality information in collaborativeﬁltering for new users,in:Proceedings of the2010ACM Conference on Recommender Systems,2010,pp.17–24.

[102]Y.Hu,Y.Koren,C.H.Volinsky,Collaborativeﬁltering for implicit feedback datasets,in:IEEE International Conference on Data Mining(ICDM),2008,pp.

263–272.

[103]Y.P.Huang,W.P.Chuang,Y.H.KE,F.E.Sandnes,Using back-propagation to learn association rules for service personalization,Expert Systems with Applications35(2008)245–253.

[104]Z.Huang, D.Zeng,H.Chen,A comparison of collaborativeﬁltering recommendation algorithms for e-commerce,IEEE Intelligent Systems22

(5)(2007)68–78.

[105]N.Hurley,M.Zhang,Novelty and diversity in top-N recommendations-analysis and evaluation,ACM Transactions on Internet Technology10(4) (2011)1–29.[106]CH.S.Hwang,Y.CH.Su,K.CH.Tseng,Using genetic algorithms for personalized recommendation,Lecture Notes in Computer Science22 (2010)104–112.

[107]H.Ingoo,J.O.Kyong,H.R.Tae,The collaborativeﬁltering recommendation based on SOM cluster-indexing CBR,Expert Systems with Applications25 (2003)413–423.

[108]A.Jameson, B.Smyth,Recommendation to groups,in:P.Brusilovsky, A.

Kobsa,W.Nejdl(Eds.),The Adaptive Web,2007,pp.596–627(Chapter20). [109]D.Jannach,Fast computation of query relaxations for knowledge-based recommenders,AI Communications22(4)(2009)235–248.

[110]R.Jäschke,L.Marinho, A.Hotho,L.Schmidt-Thieme,G.Stumme,Tag Recommendations in Folksonomies,in:Proceedings of the11th European Conference on Principles and Practice of Knowledge Discovery in Databases, 2007,pp.506–514.

[111]J.J.Jason,Contextualized mobile recommendation service based on interactive social network discovered from mobile users,Expert Systems with Applications36(9)(2009)11950–11956.

[112]B.Jeong,J.Lee,H.Cho,User credit based collaborativeﬁltering,Expert Systems with Applications36(2009)7309–7312.

[113]T.Joachims,Text categorization with support vector machines:learning with many relevant features,in:European Conference on Machine Learning,1998, pp.137–142.

[114]A.Jøsang,R.Ismail,C.Boyd,A survey of trust and reputation systems for online service provision,Decision Support Systems43(2)(2007)618–4. [115]CH.W.Jyun,CH.CH.Chui,Recommending trusted online auction sellers using social network analysis,Expert Systems with Applications34(3)(2008) 1666–1679.

[116]C.Kaleli,H.Polat,Privacy-preserving SOM-based recommendations on horizontally distributed data,Knowledge Based Systems33(2012) 124–135.

[117]H.N.Kim,A.Alkhaldi,A.E.Saddik,G.S.Jo,Collaborative user modeling with user-generated tags for social recommender Systems,Expert Systems with Applications38(7)(2011)8488–8496.

[118]H.N.Kim,A.T.Ji,I.Ha,G.S.Jo,Collaborativeﬁltering based on collaborative tagging for enhancing the quality of recommendations,Electronic Commerce Research and Applications9(1)(2010)73–83.

[119]J.Kim,B.Lee,M.Shaw,H.Chang,W.Nelson,Application of decision-tree induction techniques to personalized advertisements on internet storefronts, International Journal of Electronic Commerce5(3)(2001)45–62.

[120]K.Kim,H.Ahn,Using a clustering genetic algorithm to support customer segmentation for personalizad recommender systems,in:Proceedings of the 13th International Conference on AI,Simulation,and Planning in High Autonomy Systems,2004,pp.409-415.

[121]K.Kim,H.Ahn,A recommender system using GA K-means clustering in an online Shopping market,Expert Systems with Applications34(2)(2008) 1200–1209.

[122]S.Kitisin,C.Neuman,Reputation-based trust-aware recommender system, in:Securecomm and Workshops,2009,pp.1–7.

[123]F.Kong,X.Sun,S.Ye,A comparison of several algorithms for collaborative ﬁltering in startup stage,IEEE Transactions on Networks,Sensing and Control (2005)25–28.

[124]Y.Koren,R.Bell,CH.Volinsky,Matrix factorization techniques dor recommender systems,IEEE Computer42(8)(2009)42–49.

[125]G.Koutrika, B.Bercovitz,H.Garcia,FlexRecs:expressing and combining ﬂexible recommendations,in:Proceedings of the35th SIGMOD International Conference on Management of Data,2009,pp.745–757.

[126]B.Krulwich,Lifestyleﬁnder:intelligent user proﬁling using large-scale demographic data,Artiﬁcial Intelligence Magazine18(2)(1997) 37–45.

[127]K.Kwon,J.Cho,Y.Park,Multidimensional credibility model for neighbor selection in collaborative recommendation,Expert Systems with Applications36(2009)7114–7122.

[128]S.K.Lam,J.Riedl,Shilling recommender systems for fun and proﬁt,in: International Conference on World Wide Web,2004,pp.393–402.

[129]X.N.Lam,T.Vu,T.D.Le, A.D.Duong,Addressing cold-start problem in recommendation systems,in:Conference On Ubiquitous Information Management And Communication,2008,pp.208–211.

[130]N.Landia,S.S.Anand,Personalised tag recommendation,in:Proceedings of the2009ACM Conference on Recommender Systems,2009,pp.83–36. [131]K.Lang,NewsWeeder:learning toﬁlter netnews,in:Proceedings12th International Conference on Machine Learning,1995,pp.331–339.

[132]D.H.Lee,P.Brusilovsky,Does trust inﬂuence information similarity?in: Proceedings of the2009ACM Conference on Recommender Systems,2009, pp.71–74.

[133]M.Lee,Y.Woo,A hybrid recommender system combining collaborative ﬁltering with neural network,Lecture Notes on Computer Sciences2347 (2002)531–534.

[134]S.K.Lee,Y.H.Cho,S.H.Kim,Collaborativeﬁltering with ordinal scale-based implicit ratings for mobile music recommendations,Information Sciences 180(11)(2010)2142–2155.

[135]C.W.Leung,S.C.Chan, F.L.Chung,An empirical study of a cross-level association rule mining approach to cold-start recommendations,Knowledge Based Systems21(7)(2008)515–529.

[136]Q.Li,Clustering approach for hybrid recommender system,in:Proceedings of the2003IEEE/WIC International Conference on Web Intelligence,2003,pp.

33–38.[137]Y.M.Li,CH.W.Chen,A synthetical approach for blog recommendation: combining trust,social relation,and semantic análisis,Expert Systems with Applications36(3)(2009)6536–6547.

[138]Y.M.Li,CH.P.Kao,TREPPS:a trust-based recommender system for peer production services,Expert Systems with Applications36(2)(2009)3263–3277.

[139]Y.M.Li,T.F.Liao,CH.Y.Lai,A social recommender mechanism for improving knowledge sharing in online forums,Information Processing and Management,in press,doi:10.106/j.ipm.2011.10.004.

[140]S.Loh,F.Lorenzi,R.Granada,D.Lichtnow,L.K.Wives,J.P.Oliveira,Identifying similar users by their scientiﬁc publications to reduce cold start in recommender systems,in:Proceedings of the5th International Conference on Web Information Systems and Technologies(WEBIST2009),2009,pp.

593–600.

[141]L.Lü,M.Medo,Ch.H.Yeung,Y.Ch.Zhang,Z.K.Zhang,T.Zhou,Recommender systems,Physics Reports519(2012)1–49.

[142]X.Luo,Y.Xia,Q.Zhu,Incremental collaborativeﬁltering recommender based on regularizad matrix factorization,Knowledge-Based Systems27(2012) 271–280.

[143]X.Luo,Y.Xia,Q.Zhu,Applying the learning rate adaptation to the matrix factorization based collaborativeﬁltering,Knowledge Based Systems37 (2013)154–1.

[144]H.Ma,I.King,M.R.Lyu,Learning to recommend with explicit and implicit social relations,ACM Transactions on Intelligent Systems and Technology2

(3)(2011).Article29.

[145]H.Ma,T.CH.Zhou,M.R.Lyu,I.King,Improving recommender systems by incorporating social contextual information,ACM Transactions on Information Systems29(2)(2011).Article9.

[146]A.Machanavajjhala, A.Korolova, A.D.Sharma,Personalized social recommendations:accurate or private,in:Proceedings of the VLDB Endowment,vol.4,issue7,2011,pp.440–450.

[147]L.B.Marinho,L.Schmidt-Thieme,Collaborative tag recommendations,in: Proceedings of the31st Annual Conference of the German Classiﬁcation Society,2008,pp.533–540.

[148]L.Martinez,L.G.Perez,M.J.Barranco,Incomplete preference relations to smooth out the cold-start in collaborative recommender systems,in: Proceedings of the28th North American Fuzzy Information Processing Society Annual Conference(NAFIPS2009),2009a,pp.1–6.

[149]L.Martinez,R.M.Rodriguez,M.Espinilla,REJA:a georeferenced hybrid recommender system for restaurants,in:IEEE/WIC/ACM International Joint Conference on Web Intelligence and Intelligent Agent Technonolgy(WI-IAT

3),2009b,pp.187–190.

[150]P.Massa,P.Avesani,Trust-aware collaborativeﬁltering for recommender systems,Lecture Notes in Computer Science3290(2004)492–508.

[151]P.Massa,P.Avesani,Trust-aware recommender Systems,in:Proceedings of the2007ACM conference on Recommender Systems,2007,pp.17–24. [152]C.Matyas, C.Schlieder,A spatial user similarity measure for geographic recommender systems,in:Proceedings of the3rd International Conference on GeoSpatial Semantics,2009,pp.122–139.

[153]K.Mccarthy,J.Reilly,L.Mcginty,B.Smyth,Thinking positively-explanatory feedback for conversational recommender systems,in:European Conference on Case-based reasoning(ECCBR),2004,pp.115–124.

[154]K.Mcnally,M.P.O’mahony,M.Coyle,P.Briggs,B.Smyth,A case study of collaboration and reputation in social web search,ACM Transactions on Intelligent Systems and Technology3(1)(2011).Article4.

[155]D.Mcsherry,Explanation in recommender systems,Artiﬁcial Intelligence Review24(2)(2005)179–197.

[156]F.Mcsherry,I.Mironov,Differentially Private recommender systems: building privacy into the netﬂix prize contenders,in:Proceedings of the 15th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining(KDD),2009,pp.627–636.

[157]P.Melville,R.J.Mooney,R.Nagarajan,Content-boosted collaborativeﬁltering for improved recommendations,in:Proceeding Eighteenth National Conference on Artiﬁcial Intelligence,2002,pp.187–192.

[158]R.Meteren,M.Someren,Using content-basedﬁltering for recommendation, in:Proceedings of ECML2000Workshop:Maching Learning in Information Age,2000,pp.47–56.

[159]S.E.Middleton,N.R.Shadbolt,D.C.De Roure,Ontological user proﬁling in recommender systems,ACM Transactions on Information Systems(TOIS)22

(1)(2004)54–88.

[160]R.J.Mooney,L.Roy,Content-based book recommending using learning for text categorization,in:Proceedings of the Fifth ACM Conference on Digital Libraries,2000,pp.195–204.

[161]T.Morrison,U.Aickelin,An artiﬁcial immune system as a recommender for Web sites,in:International Conference on Artiﬁcial Immune Systems,2002, pp.161–169.

[162]A.Nanolopoulus,D.Rafailidis,P.Symeonidis,Y.Manolopoulus,Music Box: personalizad music recommendation based on cubic analysis of social tags, IEEE Transactions on Audio,Speech and Language Processing18(2)(2010) 407–412.

[163]K.Nehring,C.Puppe,A theory of diversity,Econometrica70(3)(2002)1155–1198.

[1]E.R.Núñez-Valdéz,J.M.Cueva-Lovelle,O.Sanjuán-Martı´nez,V.Garcı´a-Dı´az,P.

Ordoñez, C.E.Montenegro-Marı´n,Implicit feedback techniques on recommender systems applied to electronic books,Computers in Human Behavior28(4)(2012)1186–1193.[165]J.O’donovan,Capturing trust in social web applications,in:J.Golbeck(Ed.), Computing with Social Trust,2009,pp.213–257.

[166]J.O’donovan,B.Smyth,Trust in recommender systems,in:International Conference on Intelligent user Interfaces,2005,pp.167–174.

[167]K.Oku,R.Kotera,K.Sumiya,Geographical recommender system based on interaction between map operation and category selection,in:Workshop on Information Heterogeneity and Fusion in Recommender Systems,2010,pp.

71–74.

[168]J.Ortega,J.Bobadilla, A.Hernando, A.Gutiérrez,Incorporating group recommendations to recommender systems:alternatives and performance, Information Processing and Management(2013),http://dx.doi.org/10.1016/ j.ipm.2013.02.003.

[169]J.Ortega,J.L.Sánchez,J.Bobadilla, A.Gutiérrez,Improving collaborative ﬁltering-based recommender systems results using Pareto dominance, Information Sciences(2013),http://dx.doi.org/10.1016/j.ins.2013.03.011. [170]A.Papadimitriou,P.Symeonidid,Y.Manolopoulus,A generalized taxonomy of explanations styles for traditional and social recommender systems,Data Minning Knowledge Discovery24(3)(2012)555–583.

[171]D.H.Park,H.K.Kim,I.Y.Choi,J.K.Kim,A literature review and classiﬁcation of recommender Systems research,Expert Systems with Applications39(2012) 10059–10072.

[172]S.T.Park,W.Chu,Pairwise preference regression for cold-start recommendation,in:Proceedings of the2009ACM Conference on Recommender Systems,2009,pp.21–28.

[173]S.T.Park,D.M.Pennock,O.Madani,N.Good,D.Coste,Naı¨veﬁlterbots for robust cold-start recommendations,in:Proceedings of Knowledge Discovery and Data Mining(KDD2006),2006,pp.699–705.

[174]Y.J.Park,A.Tuzhilin,The long tail of recommender systems and how to leverage it,in:Proceedings of the2008ACM Conference on Recommender Systems,2008,pp.11–18.

[175]M.Pazzani,D.Billsus,Learning and revising user proﬁles:the identiﬁcation of interesting web sites,Machine Learning27(3)(1997)313–331.

[176]M.J.Pazzani, D.Billsus,Content-based recommender systems,in:P.

Brusilovsky,A.Kobsa,W.Nejdl(Eds.),The Adaptive Web,2007,pp.291–324(Chapter10).

[177]M.Pazzani,A framework for collaborative,content-based,and demographic ﬁltering,Artiﬁcial Intelligence Review-Special Issue on Data Mining on the Internet13(5-6)(1999)393–408.

[178]S.Perugini,M.A.Gonçalves, E.A.Fox,Recomender systems research:a connection-centric Surrey,Journal of Intelligen Information Systems23(2) (2004)107–143.

[179]M.C.Pham,Y.Cao,R.Klamma,M.Jarke,A clustering approach for collaborativeﬁltering recommendation using social network analysis, Journal of Universal Computer Science17(4)(2011)583–604.

[180]G.Pitsilis,S.J.Knapskog,Socila trust as a solution to address sparsity-inherent problems of recommender Systems,in:Proceedings of the2009ACM Conference on Recommender Systems,2009,pp.33–40.

[181]G.Pitsilis,X.Zhang,W.Wang,Clustering recommenders in collaborative ﬁltering using explicit trust information,Advances in Information and Communication Technology358(2011)82–97.

[182]A.Popescul,L.H.Ungar,D.M.Pennock,S.Lawrence,Probabilistic models for uniﬁed collaborative and content-based recommendation in sparse-data environments,in:Proceeding UAI’01Proceedings of the17th Conference in Uncertainty in Artiﬁcial Intelligence,2001,pp.437–444.

[183]C.Porcel,E.Herrera-Viedma,Dealing with incomplete information in a fuzzy linguistic recommender system to disseminate information in university digital libraries,Knowledge-Based Systems23(1)(2010)32–39.

[184]C.Porcel,J.M.Moreno,E.Herrera-Viedma,A multi-disciplinar recommender system to advice research resources in university digital libraries,Expert Systems with Applications36(10)(2009)12520–12528.

[185]C.Porcel, A.Tejeda-Lorente,M.A.Martı´nez, E.Herrera-Viedma,A hybrid recommender system for the selective dissemination of research resources in

a technology transfer ofﬁce,Information Sciences184(1)(2012)1–19. [186]J.Preece, B.Shneiderman,The reader to leader framework:motivating technology-mediated social participation,AIS Transactions on Human–Computer Interaction1(1)(2009)13–32.

[187]P.Pu,L.Chen,Trust-inspiring explanation interfaces for recommender systems,Knowledge Based Systems20(2007)542–556.

[188]W.Qin,L.Xin,H.Liang,Unifying user-based and item-based algorithm to improve collaborativeﬁltering accuracy,Energy Procedia13(2011)8231–8239.

[1]L.Ramaswamy,P.Deepak,R.Polavarapu,K.Gunasekera, D.Garg,K.

Visweswariah,S.Kalyanaraman,CAESAR:a context-aware,social recommender system for low-end mobile devices,in:International Conference on Mobile Data Management:Systems,Services and Middleware,2009,pp.338–347.

[190]A.M.Rashid,G.Karypis,J.Riedl,Learning preferences of new users in recommender systems:an information theoretic approach,in:ACM SIGKDD Explorations Newsletter,vol.10,issue2,2008,pp.90–100.

[191]S.Ray, A.Mahanti,Strategies for effective shilling attacks against recommender systems,Lecture Notes in Computer Science5456(2009) 111–125.

[192]L.Ren,L.HE,J.Gu,W.Xia,F.Wu,A hybrid recommender approach based on Widrow–Hoff learning,in:International Conference on Future Generation Communication and Networking,2008,pp.40–45.[193]T.H.Roh,K.J.Oh,I.Han,The collaborativeﬁltering recommendation based on SOM cluster-indexing CBR,Expert Systems with Applications25(2003)413–423.

[194]J.A.Rodrigues,L.F.Cardoso,J.Moreira,G.Xexeo,Bringing knowledge into recommender systems,The Journal of Systems and Software,in press,http:// dx.doi.org/10.1016/j.jss.2012.10.002.

[195]S.B.Roy,S.Amer-Yahia,A.Chala,G.Das,C.Yu,Space efﬁciency in group recommendation,The International Journal on Very Large Data Bases19(6) (2010)877–900.

[196]G.Ruffo,R.Schifanella,A peer-to-peer recommender system base don spontaneous afﬁnities,ACM Transactions on Internet Technology9(1)(2009) 1–34.

[197]P.B.Ryan, D.Bridge,Collaborative recommending using formal concept analysis,Knowledge Based Systems19(5)(2006)309–315.

[198]G.Salton,Automatic Text Processing:The Transformation,Analysis,and Retrieval of Information by Computer,Addison-Wesley,Reading,MA,19. [199]M.Saranya,T.Atsuhiro,Hybrid recommender systems using latent features, in:Proceedings of the International Conference on Advanced Information Networking and Applications Workshops,2009,pp.661–666.

[200]B.Sarwar,G.Karypis,J.A.Konstan,J.Riedl,Item-based collaborativeﬁltering recommendation algorithms,in:10th International Conference on World Wide Web,2001,pp.285–295.

[201]B.Sarwar,G.Karypis,J.Konstan,J.Riedl,Analysis of recommendation algorithms for e-commerce,in:ACM Conference on Electronic Commerce, 2000a,pp.158–167.

[202]B.Sarwar,G.Karypis,J.Konstan,J.Riedl,Application of dimensionality reduction in recommender system–a case study,in:ACM WebKDD Workshop,2000b,pp.2–272.

[203]J.B.Schafer, D.Frankowski,J.Herlocker,S.Sen,Collaborativeﬁlltering recommender systems,in:P.Brusilovsky, A.Kobsa,W.Nejdl(Eds.),The Adaptive Web,2007,pp.291–324(Chapter9).

[204]A.I.Schein,A.Popescul,L.H.Ungar,D.M.Pennock,Methods and metrics for cold-start recommendations,in:Proceeding SIGIR’02Proceedings of the25th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval,2002,pp.253–260.

[205]C.Schlieder,Modeling collaborative semantics with a geographic recommender,in:Workshop on Semantic and Conceptual Issues in Geographic Information Systems,2007,pp.336–347.

[206]J.Serrano-Guerrero,E.Herrera-Viedma,J.A.Olivas,A.Cerezo,F.P.Romero,A google wave-based fuzzy recommender system to disseminate information in University Digital Libraries2.0.,Information Sciences181(9)(2011)1503–1516.

[207]Z.Severac,V.Devedzic,J.Jovanovic,Adaptive neuro-fuzzy pedagogical recommender,Expert Systems with Applications39(10)(2012)9797–9806. [208]A.Shepitsen,J.Gemmell, B.Mobasher,R.Burke,Personalized recommendation in social tagging systems using hierarchical clustering,in: Proceedings of the2008ACM Conference on Recommender Systems,2008, pp.259–266.

[209]S.K.Shinde,U.Kulkami,Hybrid personalizad recommender system using centering–bunching based clustering algorithm,Expert Systems with Applications39(1)(2012)1381–1387.

[210]S.Siersdorfer,S.Sergei,Social recommender systems for web 2.0 folksonomies,in:20th ACM conference on Hypertext and hipermedia, 2009,pp.261–269.

[211]I.Soboroff,C.Nicholas,Combining content and collaboration in textﬁltering, in:Proceedings of the IJCAI’99Workshop on Machine Learning for Information Filtering,1999,pp.86–91.

[212]X.Su,T.M.Khoshgoftaar,A survey of collaborativeﬁltering techniques, Advance in Artiﬁcial Intelligence2009(2009)1–19.

[213]P.Symeonidis,A.Nanopoulus,Y.Manolopoulus,Providing justiﬁcations in recommender systems,IEEE Transactions on Systems,Man and Cybernet38

(6)(2008)1262–1272.

[214]P.Symeonidis, A.Nanopoulus,Y.Manolopoulus,MovieExplain:a recommender system with explanations,in:Proceedings of the2009ACM Conference on Recommender Systems,2009,pp.317–320.

[215]G.Takács,I.Pilászy, B.Németh, D.Tikk,Scalable collaborativeﬁltering approaches for large recommender systems,Journal of Machine Learning Research10(2009)623–656.

[216]S.Tan,J.Bu,CH.Chen,X.He,Using rich social media information for music recommendation via hypergraph model,ACM Transactions on Multimedia Computing,Communications,and Applications7(1)(2011).Article7. [217]N.Tintarev,J.Masthoff,A survey of explanations in recommender systems, in:IEEE23rd International Conference on Data Engineering Workshop,2007, 801–810.

[218]T.Tran,R.Cohen,Hybrid recommender systems for electronic commerce,in: Proceedings of the17th National Conference on Artiﬁcial Intelligence,AAAI, 2000,pp.78–84.

[219]K.H.L.Tso-Sutter,L.B.Marinho,L.Schmidt-Thieme,Tag-aware recommender systems by fusion of collaborativeﬁltering algorithms,in:Proceedings of the 2008ACM Symposium on Applied Computing,2008,pp.1995–1999. [220]S.Vargas,P.Castells,Rank and relevance in novelty and diversity metrics for recommender systems,in:Proceedings of the2011ACM Conference on Recommender Systems,2011,pp.109–116.

[221]P.Victor,CH.Cornelis,M.De-Cock,Trust Networks for Recommender Systems,Antalis Press,2011.[222]J.Vig,S.Sen,J.Riedle,Tagsplanations:Explaining recommendations using tags,Proceedings of the13th international conference on Intelligent user interfaces,2009,pp.47-56.

[223]P.Victor,CH.Cornelis,M.De-Cock,P.P.DA-SILVA,Gradual tust and distrust in recommender systems,Fuzzy Sets and Systems160(10)(2009)1367–1382. [224]M.G.Vozalis,K.G.Margaritis,Using SVD and demographic data for the enhancement of generalized collaborativeﬁltering,Information Sciences177 (2007)3017–3037.

[225]Y.Wan-Shiou,CH.Hung-Chi, D.Jia-Ben,A location-aware recommender system for mobile shopping environments,Expert Systems with Applications 34(1)(2008)437–445.

[226]J.Wang, A.Vries,M.Reinders,Unifying user-based and item-based collaborativeﬁltering approaches by similarity fusion,in:Proc.SIGIR Conf., 2006,pp.501–508.

[227]J.Wang, A.P.Vries,M.J.Reinders,Uniﬁed relevance models for rating prediction in collaborativeﬁltering,ACM Transactions in Information Systems26(3)(2008)1–42.

[228]L.T.Weng,Y.Xu,Y.Li,R.Nayak,Exploiting item taxonomy for solving cold-start problem in recommendation making,in:Proceedings of the20th IEEE International Conference on Tools with Artiﬁcial Intelligence(ICTAI2008), 2008,pp.113–120.

[229]B.Widrow,M.E.Hoff,Adaptive switching circuits,in:Convention Record,IRE WESCON,1960,pp.96–104.

[230]P.Winoto,T.Y.Tang,The role of user mood in movie recommendations, Expert Systems with Applications37(8)(2010)6086–6092.

[231]W.Woerndl,G.Groh,Utilizing physical and social context to improve recommender systems,in:IEEE/WIC/ACM International Conferences on Web Intelligence and Intelligent Agent Technology,2007,pp.123–128.

[232]B.Xie,P.Han,F.Yang,R.M.Shen,H.J.Zeng,Z.Chen,DCFLA:a distributed collaborative-ﬁltering neighbor-locating algorithm,Information Sciences177

(6)(2007)1349–1363.

[233]W.Xin,Q.Jamaliding,T.Okamoto,Discovering social network to improve recommender system for group learning support,in:International Conference on Computational Intelligence and Software Engineering,2009, pp.1–4.

[234]R.R.Yager,Fuzzy logic methods in recommender systems,Fuzzy Sets and Systems136(2)(2003)133–149.

[235]W.S.Yang,H.CH.Cheng,J.B.Dia,A location-aware recommender system for mobile shopping environments,Expert Systems With Applications34(1) (2008)437–445.

[236]Y.Yang,An evaluation of statistical approaches to text categorization, Information Retrieval1(1)(1999)67–88.

[237]Z.Yao,Q.Zhang,Item-based clustering collaborativeﬁltering algorithm under high dimensional sparse data,in:International Joint Confeence on Computational Sciences and Optimization,2009,pp.787–790.[238]Z.Yu,X.Zhou,Y.Hao,J.Gu,TV program recommendation for multiple viewers based on user proﬁle merging,User Modeling and User-Adapted Interaction16(1)(2006)63–82.

[239]W.Yuan, D.Guan,Y.K.Lee,S.Lee,S.J.Hur,Improved trust-aware recommender system using small-worldness of trust networks,Knowledge Based Systems23(3)(2010)232–238.

[240]G.Zacharia,A.Moukas,P.Maes,Collaborative reputation mechanisms for electronic marketplaces,Decision Support Systems29(2000)371–388. [241]O.Zaiane,Building a recommender agent for e-learning systems,in: Proceedings of the International Conference on Computers Education (ICCE’02),vol.1,2002,pp.55–59.

[242]J.Zhan,Privacy-preserving collaborative recommender systems,IEEE Transactions on Systems,Man and Cybernetics40(4)(2010)472–476. [243]F.Zhang,H.Y.Chang,A collaborativeﬁltering algorithm employing genetic clustering to ameliorate the scalability issue,in:IEEE International Conference on e-Business Engineering,2006,pp.331–338.

[244]S.Zhang,W.Wang,J.Ford,F.Makedon,Using singular value decomposition approximation for collaborativeﬁltering,in:IEEE International Conference on E-Commerce Technology,2005,pp.1–8.

[245]L.Zhen,G.Q.Huang,Z.Jiang,Recommender systems based on workﬂow, Decision Support Systems48(2009)237–245.

[246]L.Zhen,G.Q.Huang,Z.Jiang,Collaborativeﬁltering based on workﬂow space, Expert Systems with Applications36(2009)7873–7881.

[247]L.Zhen,Z.Jiang,H.Song,Distributed recommender for peer-to-peer knowledge sharing,Information Sciences210(2010)3546–3561.

[248]N.Zheng,Q.Li,A recommender system based on tag and time information for social tagging systems,Expert Systems with Applications38(4)(2011) 4575–4587.

[249]Y.Zheng,X.Xie,Learning travel recommendations from user-generated GPS traces,ACM Transactions on Intelligent Systems and Technology2(2011)1.

Article2.

[250]Y.Zheng,L.Zhang,Z.Ma,X.Xie,W.Y.Ma,Recommending friends and locations based on individual location history,ACM Transactions on the Web 5(2011)1.Article5.

[251]J.Zhong,X.Li,Uniﬁed collaborativeﬁltering model based on combination of latent features,Expert Systems with Applications37(2010)5666–5672. [252]R.L.Zhu,S.J.Gong,Analyzing of collaborativeﬁltering using clustering technology,international colloquium on computing,in:ISECS International Colloquium on Computing,Communication,Control,and Management,2009, pp.57–59.

[253]C.N.Ziegler,S.M.Mcnee,J.A.Konstan,G.Lausen,Improving recommendation lists through topic diversiﬁcation,in:Proceedings of the14th International Conference on World Wide Web,2005,pp.22–32.下载本文

显示全文

全部频道