Knowledge-Based Systems 23(2010)520-528 Contents lists available at Science Direct Knowledge-Based Systems ELSEVIER journalhomepagewww.elsevier.com/locate/knosys A new collaborative filtering metric that improves the behavior of recommender systems J. Bobadilla, F. Serradilla,J. Bernal Universidad Politecnica de madrid, Computer Science, Crta de valencia, Km 7, 28031 Madrid, Spain ARTICLE INFO A BSTRACT Recommender systems are typically provided as Web 2.0 services and are part of the range of applica- 21July2009 tions that give support to large-scale social networks, enabling on-line recommendations to be made revised form 18 March 2010 19 March 2010 ased on the use of networked databases. The operating core of recommender systems is based on the nline 23 March 2010 collaborative filtering stage, which, in current user to user recommender processes, usually uses the Pear son correlation metric. In this paper, we present a new metric which combines the numerical information of the votes with independent information from those values, based on the proportions of the common orative fltering and uncommon votes between each pair of users. Likewise, we define the reasoning and experiments on amender systems hich the design of the metric is based and the restriction of being applied to recommender where the possible range of votes is not greater than 5. In order to demonstrate the supe the proposed metric, we provide the comparative results of a set of experiments based on the Mo lean squared differences FilmAffinity and Net Flix databases. In addition to the traditional levels of accuracy, results are also pro- vided on the metrics'coverage, the percentage of hits obtained and the precision/recall 2010 Elsevier B V. All rights reserved. 1 Introduction Memory-based methods [22, 37, 35, 40 use similarity metrics and act directly on the ratio matrix that contains the ratings of Recommender systems(RS)cover an important field within col- all users who have expressed their preferences on the collaborative laborative services that are developed in the Web 2.0 environment service; these metrics mathematically express a distance between [21, 19, 26 and enable user-generated opinions to be exploited in a two users based on each of their ratios. Model-based methods [1] sophisticated and powerful way. Recommender systems can be use the ratio matrix to create a model from which the sets of sim- considered as social networking tools that provide dynamic and ilar users will be established. Among the most widely-used models collaborative communication, interaction and knowledge. we have: bayesian classifiers [8, neural networks [18 and fuzzy Recommender systems cover a wide variety of applications systems [39]. Generally, commercial recommender systems use 20, 4, 10, 28.5, 14, 27, although those related to movie recommen- memory-based methods, whilst model-based methods are usually dations are the most well-known and most widely-used in the associated with research recommender systems. search field [23, 3, 25] Regardless of the method used in the collaborative filtering filtering(CF)phase [1, 17,6, 34, 33]. Collaborative filtering is based ommender systems as high as possible: nevertheless, there are on making predictions about a users preferences or tastes based other objectives that need to be taken into account [38 avoid on the preferences of a group of users that are considered similar overspecialization phenomena, find good items, credibility of rec- to this user. A substantial part of the research in the area of collab- ommendations, precision and recall measures, etc. orative filtering centers on how to determine which users are sim- To date, various publications have been written which tackle the ilar to the given one: in order to tackle this task, there are way the recommender systems are evaluated, among the most sig- fundamentally 3 approaches: memory-based methods, model- nificant we have[ 17 which reviews the key decisions in evaluating based methods and hybrid approaches collaborative filtering recommender systems: the user tasks, the type of analysis and datasets being used, the ways in which predic- tion quality is measured and the user-based evaluation of the syster as a whole. Hernandez and gaudioso 9] is a current study which ations: RS, recommender systems: CF, collaborative filte ponding author. Tel: +34 913365133: fax: +34 913367522. proposes a recommendation filtering process based on the distinc address: jesus. bobadilla@upmes 0. Bobadilla). tion between interactive and non-interactive subsystems. General 0950-7051/s-see front matter o 2010 Elsevier B v. All rights reserved o:10.1016/ knosys2010.03009A new collaborative filtering metric that improves the behavior of recommender systems J. Bobadilla *, F. Serradilla, J. Bernal Universidad Politécnica de Madrid, Computer Science, Crta. de Valencia, Km 7, 28031 Madrid, Spain article info Article history: Received 21 July 2009 Received in revised form 18 March 2010 Accepted 19 March 2010 Available online 23 March 2010 Keywords: Collaborative filtering Recommender systems Metric Jaccard Mean squared differences Similarity abstract Recommender systems are typically provided as Web 2.0 services and are part of the range of applica￾tions that give support to large-scale social networks, enabling on-line recommendations to be made based on the use of networked databases. The operating core of recommender systems is based on the collaborative filtering stage, which, in current user to user recommender processes, usually uses the Pear￾son correlation metric. In this paper, we present a new metric which combines the numerical information of the votes with independent information from those values, based on the proportions of the common and uncommon votes between each pair of users. Likewise, we define the reasoning and experiments on which the design of the metric is based and the restriction of being applied to recommender systems where the possible range of votes is not greater than 5. In order to demonstrate the superior nature of the proposed metric, we provide the comparative results of a set of experiments based on the MovieLens, FilmAffinity and NetFlix databases. In addition to the traditional levels of accuracy, results are also pro￾vided on the metrics’ coverage, the percentage of hits obtained and the precision/recall. 2010 Elsevier B.V. All rights reserved. 1. Introduction Recommender systems (RS) cover an important field within col￾laborative services that are developed in the Web 2.0 environment [21,19,26] and enable user-generated opinions to be exploited in a sophisticated and powerful way. Recommender systems can be considered as social networking tools that provide dynamic and collaborative communication, interaction and knowledge. Recommender systems cover a wide variety of applications [20,4,10,28,5,14,27], although those related to movie recommen￾dations are the most well-known and most widely-used in the re￾search field [23,3,25]. The recommender systems stage that normally has the greatest influence on the quality of the results obtained is the collaborative filtering (CF) phase [1,17,6,34,33]. Collaborative filtering is based on making predictions about a user’s preferences or tastes based on the preferences of a group of users that are considered similar to this user. A substantial part of the research in the area of collab￾orative filtering centers on how to determine which users are sim￾ilar to the given one; in order to tackle this task, there are fundamentally 3 approaches: memory-based methods, model￾based methods and hybrid approaches. Memory-based methods [22,37,35,40] use similarity metrics and act directly on the ratio matrix that contains the ratings of all users who have expressed their preferences on the collaborative service; these metrics mathematically express a distance between two users based on each of their ratios. Model-based methods [1] use the ratio matrix to create a model from which the sets of sim￾ilar users will be established. Among the most widely-used models we have: bayesian classifiers [8], neural networks [18] and fuzzy systems [39]. Generally, commercial recommender systems use memory-based methods, whilst model-based methods are usually associated with research recommender systems. Regardless of the method used in the collaborative filtering stage, the technical objective generally pursued is to minimize the prediction errors, by making the accuracy [12,11,29] of the rec￾ommender systems as high as possible; nevertheless, there are other objectives that need to be taken into account [38]: avoid overspecialization phenomena, find good items, credibility of rec￾ommendations, precision and recall measures, etc. To date, various publications have been written which tackle the way the recommender systems are evaluated, among the most sig￾nificant we have [17] which reviews the key decisions in evaluating collaborative filtering recommender systems: the user tasks, the type of analysis and datasets being used, the ways in which predic￾tion quality is measured and the user-based evaluation of the system as a whole. Hernández and Gaudioso [9] is a current study which proposes a recommendation filtering process based on the distinc￾tion between interactive and non-interactive subsystems. General 0950-7051/$ - see front matter 2010 Elsevier B.V. All rights reserved. doi:10.1016/j.knosys.2010.03.009 Abbreviations: RS, recommender systems; CF, collaborative filtering. * Corresponding author. Tel.: +34 913365133; fax: +34 913367522. E-mail address: jesus.bobadilla@upm.es (J. Bobadilla). Knowledge-Based Systems 23 (2010) 520–528 Contents lists available at ScienceDirect Knowledge-Based Systems journal homepage: www.elsevier.com/locate/knosys