PERUGINI, GONCALVES AND FOX text-intensive domains, which account for only a portion of the artifact landscape. Since we take a connection-oriented perspective toward recommendation, content-based models and methods do not find place in this survey In addition to identifying these differences, articles in this special issue also reported new research developments. Foltz and Dumais introduced latent semantic indexing as a viable technique to reduce dimensions in a term-document matrix(Foltz and Dumais, 1992). More importantly for recommender systems, Goldberg et al. coined the phrase collaborative filtering( Goldberg et al, 1992) while describing Tapestry, which later ame known as the first recommender system(Resnick and Varian, 1997). Collabora tive filtering, which can be defined as harnessing the activities of others in satisfying an information-seeking goal, introduced another shift in IS research. Collaborative filtering entails filtering items for a user that similar users filtered. Instead of computing artifact similarity(content-based filtering), collaborative approaches entail computing user simi larity. The most salient difference between these two approaches is that in content-based filtering users do not collaborate to improve the systems model of them, while in col- laborative approaches users leverage the collective experience of other users to enrich the systems model. Collaborative filtering is predicated upon persistent user models, such as profiles, which encapsulate preferences and features(e.g, married), rather than ephemer queries. This shift replaced features with representations of people(e.g, rating or profiles)to filter documents in a modeling matrix. While documents still constituted the other dimen- sion of the matrix, the word'document assumed a broader meaning after the birth of the web. In addition to its traditional interpretation, it also came to mean webpages and bookmarks(Balabanovic and Shoham, 1997: Rcuker and polano, 1997: Terveen et al 1997), as well as Usenet and e-mail messages( Goldberg et al., 1992; Konstan et al. Collaborative-filtering is effective since peoples tastes are typically not orthogonal How- ever, initially it was not embraced. Meanwhile, the advent of the web and its widespread use popularity, and acceptance, made reducing information overload a necessity. Of particular importance was social information filtering, a concept developed by Shardanand and Maes (1995). A few years later, in 1996, interest in collaborative filtering led to a workshop on the topic at the University of California, Berkeley. The results of this Berkeley workshop led to the March 1997 Communications of the ACM special issue on recommender systems, a phrase coined by Resnick and Varian in their article introducing the issue(Resnick and Varian, 1997) Resnick and Varian choose the phrase recommender systems'rather than'collabora- tive filtering because recommenders need not explicitly collaborate with recommendation recipients, if at all(helping to reconcile the differences between content-based and collab- orative approaches)(Resnick and Varian, 1997). Furthermore, recommendation refers to suggesting interesting artifacts in addition to solely filtering undesired objects(helping to reconcile the differences between IR and IF). Resnick and Varian define a recommender as a system which accepts user models as input, aggregates them, and returns recommen- dations to users. Two early collaborative-filtering recommender systems were Firefly and LikeMinds. Firefly evolved from Ringo(Shardanand and Maes, 1995)and HOMR (Helpful
112 PERUGINI, GONC¸ ALVES AND FOX text-intensive domains, which account for only a portion of the artifact landscape. Since we take a connection-oriented perspective toward recommendation, content-based models and methods do not find place in this survey. In addition to identifying these differences, articles in this special issue also reported new research developments. Foltz and Dumais introduced latent semantic indexing as a viable technique to reduce dimensions in a term-document matrix (Foltz and Dumais, 1992). More importantly for recommender systems, Goldberg et al. coined the phrase collaborative filtering (Goldberg et al., 1992) while describing Tapestry, which later became known as the first recommender system (Resnick and Varian, 1997). Collaborative filtering, which can be defined as harnessing the activities of others in satisfying an information-seeking goal, introduced another shift in IS research. Collaborative filtering entails filtering items for a user that similar users filtered. Instead of computing artifact similarity (content-based filtering), collaborative approaches entail computing user similarity. The most salient difference between these two approaches is that in content-based filtering users do not collaborate to improve the system’s model of them, while in collaborative approaches users leverage the collective experience of other users to enrich the system’s model. Collaborative filtering is predicated upon persistent user models, such as profiles, which encapsulate preferences and features (e.g., married), rather than ephemeral queries. This shift replaced features with representations of people (e.g., rating or profiles) to filter documents in a modeling matrix. While documents still constituted the other dimension of the matrix, the word ‘document’ assumed a broader meaning after the birth of the web. In addition to its traditional interpretation, it also came to mean webpages and bookmarks (Balabanovi´c and Shoham, 1997; Rcuker and Polano, 1997; Terveen et al., 1997), as well as Usenet and e-mail messages (Goldberg et al., 1992; Konstan et al., 1997). Collaborative-filtering is effective since people’s tastes are typically not orthogonal. However, initially it was not embraced. Meanwhile, the advent of the web and its widespread use, popularity, and acceptance, made reducing information overload a necessity. Of particular importance was social information filtering, a concept developed by Shardanand and Maes (1995). A few years later, in 1996, interest in collaborative filtering led to a workshop on the topic at the University of California, Berkeley. The results of this Berkeley workshop led to the March 1997 Communications of the ACM special issue on recommender systems, a phrase coined by Resnick and Varian in their article introducing the issue (Resnick and Varian, 1997). Resnick and Varian choose the phrase ‘recommender systems’ rather than ‘collaborative filtering’ because recommenders need not explicitly collaborate with recommendation recipients, if at all (helping to reconcile the differences between content-based and collaborative approaches) (Resnick and Varian, 1997). Furthermore, recommendation refers to suggesting interesting artifacts in addition to solely filtering undesired objects (helping to reconcile the differences between IR and IF). Resnick and Varian define a recommender as a system which accepts user models as input, aggregates them, and returns recommendations to users. Two early collaborative-filtering recommender systems were Firefly and LikeMinds. Firefly evolved from Ringo (Shardanand and Maes, 1995) and HOMR (Helpful
RECOMMENDER SYSTEMS RESEARCH 113 Online Music Recommendation Service)and allows a website to make intelligent book. movie, or music recommendations. Fireflys underlying algorithm(Shardanand, 1994)is nowusedtopowertherecommendationenginesofsitessuchasBarnesandnoble.com Collaborative approaches constitute the main thrust of current recommender systems research. Once users are modeled, the process of collaborative filtering can be viewed oper- ationally as a function which accepts a representation of users and universal set of artifacts as input and returns a recommended subset of those artifacts as output. More importantl for this survey, recommender systems also are intended to connect groups of individuals with similar interests and to leverage the collective experience rather than merely focus- ing on the information-seeking goal of a specific individual (as in a typical IR setting) In order to make connections, this function typically computes similarity(e. g, closeness, distance, or nearest neighbor). Making recommendations and thus connections then entails approximating this function. Approaches to this approximation that have evolved range from statistical models(e. g, correlating user ratings(Konstan et al., 1997)or reducing dimensions( Goldberg et al., 2000))to attribute-value based learning techniques(e.g, deci- sion trees, neural networks, and Bayesian classifiers)(Russell and Norvig, 1995) and have demonstrated qualified success(Breese et al., 1998). Ultimately these techniques can be viewed as ways to infer structure and induce connections in the modeling matrix spac This final shift replaced documents with artifacts in the modeling matrix. While the evolution of recommender systems research is characterized by the shifts in matrix models illustrated in Table l, the sparsity and anti-symmetric properties remained constant across each. As shown below, the web makes the matrix model symmetric. Sparsity is mostly attributable to the reluctance of users to rate artifacts Reluctance results from a lack of time, patience, or willingness to participate. Sometimes the benefits gained from providin constructive feedback are not apparent initially. Reluctance may be partially attributable to a heightened awareness of privacy when divulging personal information. Therefore, collaborative-based recommender systems must mediate an accuracy(of connection)vS. sparsity(of model) tradeoff. The following two sections are devoted to strategies for filling in cells of the initially sparse modeling matrix Since 1997 recommender systems research has advanced in many directions, such as reputation systems(Resnick et al., 2000)(e.g, eBay. com), and was placed in a larger context called'personalization'(Riecken, 2000). The functional-emphasis of current recommender systems makes them templates for personalization( Perugini and Ramakrishnan, 2003) 3. Creating connections: Explicit user modeling User modeling entails developing representations of user needs, interests, and taste and is a critical precursor to connecting people via recommendation. In addition to personal char acteristics, users can be modeled by their assessments of products in the form of ratings, which then become matrix entries. Sparse user feedback is the single greatest bottleneck of any collaborative-filtering algorithm: Collaborative filtering algorithms are not deemed universally acceptable precisely because users are not willing to invest much time or effort in rating the items(Aggarwal et aL., 1999). These problems are compounded in volumi- nous domains, where a large cumulative number of ratings is required to sufficiently cover
RECOMMENDER SYSTEMS RESEARCH 113 Online Music Recommendation Service) and allows a website to make intelligent book, movie, or music recommendations. Firefly’s underlying algorithm (Shardanand, 1994) is now used to power the recommendation engines of sites such as BarnesandNoble.com. Collaborative approaches constitute the main thrust of current recommender systems research. Once users are modeled, the process of collaborative filtering can be viewed operationally as a function which accepts a representation of users and universal set of artifacts as input and returns a recommended subset of those artifacts as output. More importantly for this survey, recommender systems also are intended to connect groups of individuals with similar interests and to leverage the collective experience rather than merely focusing on the information-seeking goal of a specific individual (as in a typical IR setting). In order to make connections, this function typically computes similarity (e.g., closeness, distance, or nearest neighbor). Making recommendations and thus connections then entails approximating this function. Approaches to this approximation that have evolved range from statistical models (e.g., correlating user ratings (Konstan et al., 1997) or reducing dimensions (Goldberg et al., 2000)) to attribute-value based learning techniques (e.g., decision trees, neural networks, and Bayesian classifiers) (Russell and Norvig, 1995) and have demonstrated qualified success (Breese et al., 1998). Ultimately these techniques can be viewed as ways to infer structure and induce connections in the modeling matrix space. This final shift replaced documents with artifacts in the modeling matrix. While the evolution of recommender systems research is characterized by the shifts in matrix models illustrated in Table 1, the sparsity and anti-symmetric properties remained constant across each. As shown below, the web makes the matrix model symmetric. Sparsity is mostly attributable to the reluctance of users to rate artifacts. Reluctance results from a lack of time, patience, or willingness to participate. Sometimes the benefits gained from providing constructive feedback are not apparent initially. Reluctance may be partially attributable to a heightened awareness of privacy when divulging personal information. Therefore, collaborative-based recommender systems must mediate an accuracy (of connection) vs. sparsity (of model) tradeoff. The following two sections are devoted to strategies for filling in cells of the initially sparse modeling matrix. Since 1997 recommender systems research has advanced in many directions, such as reputation systems (Resnick et al., 2000) (e.g., eBay.com), and was placed in a larger context called ‘personalization’ (Riecken, 2000). The functional-emphasis of current recommender systems makes them ‘templates for personalization’ (Perugini and Ramakrishnan, 2003). 3. Creating connections: Explicit user modeling User modeling entails developing representations of user needs, interests, and taste and is a critical precursor to connecting people via recommendation. In addition to personal characteristics, users can be modeled by their assessments of products in the form of ratings, which then become matrix entries. Sparse user feedback is the single greatest bottleneck of any collaborative-filtering algorithm: ‘Collaborative filtering algorithms are not deemed universally acceptable precisely because users are not willing to invest much time or effort in rating the items’ (Aggarwal et al., 1999). These problems are compounded in voluminous domains, where a large cumulative number of ratings is required to sufficiently cover
114 PERUGINI, GONCALVES AND FOX an entire set of items. Moreover, as the number of dimensions(e. g, people or products) grows larger, the number of multidimensional comparisons grows. In such situations tech- from data sing and OLAP(On-Line Analytical Processing )are appl Adomavicius and Tuzhilin, 2001). In large domains, users typically examine and evaluate only a small percentage of all items. Shallow analysis of content makes fostering connec tions difficult since opportunity for user overlap is limited. While in the initial stages of a system, this challenge has been echoed as the'cold-start'problem(Maltz and Ehrlich, 1995) as the day lifetime of a system. For example, a collaborative recommender has no platform to compute connections for a new user who has yet to rate products or a new item which has yet to be evaluated. Such problems in developing a basis for collaboration provide ample motivation for hybrid approaches which employ content-based filtering in these specific situations. Hy brid systems have shown improved performance over either single focus(pure)approach (Baudisch, 1999: Claypool et al., 1999; Soboroff and Nicholas, 1999). Systems must collect user data which affords the identification of differences, commonalities, and relationship mong people. In short, the goal is to add more and more information to transform a sparse matrix to a dense matrix with added structure Approaches to user modeling can be studied by how they harvest data(Resnick and arian,1997), either explicitly by asking users to submit feedback through surveys(Konstan et al., 1997)or inferring user interest implicit in(usage)data( Claypool et al., 2001; Tervee et al., 1997).Strategies for the former approach are showcased in this section, while those for the later are discussed in Section 4. The most important tradeoff to consider in user modeling is minimizing usereffort while maximizing the expressiveness of the representation (as well as privacy). In other words, there should be a small learning curve. Explicit approaches allow the user to retain control over the amount of personal information supplied to the system, but require an investment in time and effort to yield connections. other hand, minimize effort collect copious amounts of (sometimes noisy)data, and make the social element to recommender systems salient, but raise ethical issues. The secretive nature of these approaches often make users feel as if they are under a microscope. The user-modeling methodology for a collaborative-based system is illustrated in Table 2. In explicit user modeling, evaluations(Konstan et al., 1997) and profiles( Balabanovic and Shoham, 1997) are provided directly by users to declare preferences in response to elicitations for data such as surveys. Evaluations of recommended artifacts can be both Table 2. User modeling methodology of a collaborative-filtering recommender system. reluctance to rate items(compounded by volume concern of privacy) sparse modeling matrix(cold-start) explicit implicit user modeling(exploration) representation of user(ratings, profiles)as basis for connection deliver recommendations create connections(exploitation)
114 PERUGINI, GONC¸ ALVES AND FOX an entire set of items. Moreover, as the number of dimensions (e.g., people or products) grows larger, the number of multidimensional comparisons grows. In such situations techniques from data warehousing and OLAP (On-Line Analytical Processing) are applicable (Adomavicius and Tuzhilin, 2001). In large domains, users typically examine and evaluate only a small percentage of all items. Shallow analysis of content makes fostering connections difficult since opportunity for user overlap is limited. While in the initial stages of a system, this challenge has been echoed as the ‘cold-start’ problem (Maltz and Ehrlich, 1995) (also referred to as the ‘day-one’ or ‘early-rater’ problem), it is also ubiquitous during the lifetime of a system. For example, a collaborative recommender has no platform to compute connections for a new user who has yet to rate products or a new item which has yet to be evaluated. Such problems in developing a basis for collaboration provide ample motivation for hybrid approaches which employ content-based filtering in these specific situations. Hybrid systems have shown improved performance over either single focus (pure) approach (Baudisch, 1999; Claypool et al., 1999; Soboroff and Nicholas, 1999). Systems must collect user data which affords the identification of differences, commonalities, and relationships among people. In short, the goal is to add more and more information to transform a sparse matrix to a dense matrix with added structure. Approaches to user modeling can be studied by how they harvest data (Resnick and Varian, 1997), either explicitly by asking users to submit feedback through surveys (Konstan et al., 1997) or inferring user interest implicit in (usage) data (Claypool et al., 2001; Terveen et al., 1997). Strategies for the former approach are showcased in this section, while those for the later are discussed in Section 4. The most important tradeoff to consider in user modeling is minimizing user effort while maximizing the expressiveness of the representation (as well as privacy). In other words, there should be a small learning curve. Explicit approaches allow the user to retain control over the amount of personal information supplied to the system, but require an investment in time and effort to yield connections. Implicit approaches, on the other hand, minimize effort, collect copious amounts of (sometimes noisy) data, and make the social element to recommender systems salient, but raise ethical issues. The secretive nature of these approaches often make users feel as if they are under a microscope. The user-modeling methodology for a collaborative-based system is illustrated in Table 2. In explicit user modeling, evaluations (Konstan et al., 1997) and profiles (Balabanovi´c and Shoham, 1997) are provided directly by users to declare preferences in response to solicitations for data such as surveys. Evaluations of recommended artifacts can be both Table 2. User modeling methodology of a collaborative-filtering recommender system. user reluctance to rate items (compounded by volume & concern of privacy) ↓ sparse modeling matrix (cold-start) ↓ −→ explicit + implicit user modeling (exploration) ↓ representation of user (ratings, profiles) as basis for connection ↓ −→ −→ ←− deliver recommendations & create connections (exploitation) sustain (exploration vs. exploitation)
RECOMMENDER SYSTEMS RESEARCH 115 quantitative(e. g, ratings), akin to relevance feedback in IR and IF(Mostafa et al., 1997), and qualitative(e.g, lengthy reviews at Epinions. com). They also can be positive or negative. In a hand-crafted profile, a user states interests through items such as lists of keywords, pre-defined categories, or descriptions. The system then matches other users against this profile to recommend incoming artifacts. Systems which take such an approach to user modeling are SIFT (Yan and Garcia-Molina, 1999)and Tapestry( Goldberg et al., 1992) without crossing over to an implicit approach, researchers have identified strategies to deal with reluctance to make an explicit feedback requirement less noticeable and taxing (Konstan et al., 1997; Resnick and Varian, 1997). Possible approaches to motivate users to evaluate items are subscription services, incentives, such as transaction-based compensa- tions, and exclusions(Avery and Zeckhauser, 1997). Employing a pay-per-use model for recommender systems, where human experts rate items, is a viable, though less dynamic, option. While this approach connects users through experts and is thus collaborative, it deemphasizes the naturally social (and personal)element to recommenders. Default votes are another way to deal with sparse ratings( Breese et al., 1998). Developing and tightly integrating natural user interface (Un) mechanisms to solicit and capture feedback with existing interfaces for recommendation delivery may lead to less intrusive interaction and thus more cooperation and data( Grasso et al., 1999). A similar approach is to build recom- mendation into everyday systems, such as e-mail, news, and web clients, and services like collaborativespamdetectors(e.g,Cloudmark'sSpamnet,http://www.cloudmark.com).In addition to helping to collect more explicit ratings, building recommendation into com- mon Uls may help disseminate recommender systems to the masses. Requiring users to evaluate clusters of, rather than individual, items is another approach to mini effort. Rather than tackling sparsity from a user perspective in an explicit approach, it also can be approached from a system viewpoint. Filter-bots which automatically exam- ine and rate all products may occupy empty cells of a modeling matrix(Sarwar et al Lastly, a problem endemic to the subjective nature of explicit modeling techniques is that some users are more effusive in their ratings than others. Effusivity in ratings refers to cases of users who share similar preferences, but rate products on completely different scales Identifying variations in rating patterns is an approach to combat effusivity(Aggarwal et al 1999: Fruend et al., 1998). Other considerations. A variety of representations have been used to store user data (Bloedorn et al., 1996). The lack of standards to represent such information and its sources (e. g, logs) in a uniform manner make interoperability among recommender systems a challenge(Basu and Hirsh, 2001; Cingil et al., 2000). Cookies are mechanisms for capturing and storing userpreferences, often employed in e-commerce(Berghel, 2001). While cookies combat the stateless Http protocol like many of these techniques they raise security and privacy concerns because they are typically unknowingly enabled and as a result personal information is divulged. a challenge for any user modeling approach(explicit or implicit, for content-based or ollaborative recommendation) is the tradeoff between exploration(modeling the user) and exploitation(using the model to predict future ratings or make recommendations and
RECOMMENDER SYSTEMS RESEARCH 115 quantitative (e.g., ratings), akin to relevance feedback in IR and IF (Mostafa et al., 1997), and qualitative (e.g., lengthy reviews at Epinions.com). They also can be positive or negative. In a hand-crafted profile, a user states interests through items such as lists of keywords, pre-defined categories, or descriptions. The system then matches other users against this profile to recommend incoming artifacts. Systems which take such an approach to user modeling are SIFT (Yan and Garc´ıa-Molina, 1999) and Tapestry (Goldberg et al., 1992). Without crossing over to an implicit approach, researchers have identified strategies to deal with reluctance to make an explicit feedback requirement less noticeable and taxing (Konstan et al., 1997; Resnick and Varian, 1997). Possible approaches to motivate users to evaluate items are subscription services, incentives, such as transaction-based compensations, and exclusions (Avery and Zeckhauser, 1997). Employing a pay-per-use model for recommender systems, where human experts rate items, is a viable, though less dynamic, option. While this approach connects users through experts and is thus collaborative, it deemphasizes the naturally social (and personal) element to recommenders. Default votes are another way to deal with sparse ratings (Breese et al., 1998). Developing and tightly integrating natural user interface (UI) mechanisms to solicit and capture feedback with existing interfaces for recommendation delivery may lead to less intrusive interaction and thus more cooperation and data (Grasso et al., 1999). A similar approach is to build recommendation into everyday systems, such as e-mail, news, and web clients, and services like collaborative spam detectors (e.g., Cloudmark’s SpamNet, http://www.cloudmark.com). In addition to helping to collect more explicit ratings, building recommendation into common UIs may help disseminate recommender systems to the masses. Requiring users to evaluate clusters of, rather than individual, items is another approach to minimizing effort. Rather than tackling sparsity from a user perspective in an explicit approach, it also can be approached from a system viewpoint. Filter-bots which automatically examine and rate all products may occupy empty cells of a modeling matrix (Sarwar et al., 1997). Lastly, a problem endemic to the subjective nature of explicit modeling techniques is that some users are more effusive in their ratings than others. Effusivity in ratings refers to cases of users who share similar preferences, but rate products on completely different scales. Identifying variations in rating patterns is an approach to combat effusivity (Aggarwal et al., 1999; Fruend et al., 1998). Other considerations. A variety of representations have been used to store user data (Bloedorn et al., 1996). The lack of standards to represent such information and its sources (e.g., logs) in a uniform manner make interoperability among recommender systems a challenge (Basu and Hirsh, 2001; Cingil et al., 2000). Cookies are mechanisms for capturing and storing user preferences, often employed in e-commerce (Berghel, 2001). While cookies combat the stateless HTTP protocol, like many of these techniques, they raise security and privacy concerns because they are typically unknowingly enabled and as a result personal information is divulged. A challenge for any user modeling approach (explicit or implicit, for content-based or collaborative recommendation) is the tradeoff between exploration (modeling the user) and exploitation (using the model to predict future ratings or make recommendations and
116 PERUGINI, GONCALVES AND FOX connections), akin to that in reinforcement learning(Sutton and Barto, 1998). Studying the connections which can be made via recommendation and the resulting social networl induced in a random graph setting provides technical insight into this problem. Mirza et al. (2003)identify a'minimum rating constraint required to sustain a system and predict values for it based on various experimental rating datasets. Ultimately the approaches to user modeling illustrated in this and the following section are used to connect people. While a purely collaborative approach to recommendation is widely accepted and employed, it is riddled with endemic problems. User modeling must address more than just sparsity. For example, it is difficult to make connections to users with unusual or highly specific tastes. Furthermore, connecting users with similar interests who have rated different items(e.g, we both read world politics online, but he ranked BBC. com webpages, while I ranked CNN. com pages)is challenging. Over-specialization of evaluated artifacts, sometimes referred to as the banana problem(Burke, 1999), arises since frequently purchased items, such as bananas in a grocery market basket, will al ways be recommended. Conversely, some products are seldomly bought more than a few times in a lifetime(e. g, automobiles) and thus suffer from a low number of evaluations Over-specialization which is grounded in the exploration vs. exploitation dilemma can be addressed by occasionally forcing exploration. For instance, one can inject random- ess(e.g, crossover and mutation in a genetic algorithm or epsilon in a reinforcement learning algorithm) into a model. Recommended artifacts also can be partitioned into hot and cold sets, where the latter is intended to foster exploration and increase the(rating) coverage of items in the system(Aggarwal et al., 1999) 3.1. Review of some representative projects The following collaborative-based systems employ many of the explicit user modeling techniques showcased above and illustrate what can be achieved with representations of users. People are connected in the following systems through statistical( Goldberg et al., 2000: Konstan et al., 1997), agent-oriented (Balabanovic and Shoham, 1997), and graph theoretic(Aggarwal et al., 1999)approache Group Lens. GroupLens recommends Usenet news messages(Konstan et al., 1997). The system models users directly by explicitly eliciting and collecting ratings of messages through an independent newsreader. GroupLens is a project of the recommender systems research group at the University of Minnesota. Usenet news is a personal, voluminous, and ephemeral media(in comparison to movies)and thus an excellent candidate for collaborative filtering. A total of 250 people evaluated over 20,000 news articles(Resnick et al., 1994) GroupLens takes a statistical approach to making connections. The system predicts how a user seeking recommendation would rate an unrated article by computing a weighted average of the ratings of that message by users whose ratings were correlated with the user seeking recommendation Correlation is computed with Pearsons r coefficient. A research issue is deciding whether to provide personalized predictions(as GroupLens currently does) vS. personalized averages. Empirical research using Pearsons r correla- tion coefficient revealed that correlations between ratings and predictions is dramatically
116 PERUGINI, GONC¸ ALVES AND FOX connections), akin to that in reinforcement learning (Sutton and Barto, 1998). Studying the connections which can be made via recommendation and the resulting social network induced in a random graph setting provides technical insight into this problem. Mirza et al. (2003) identify a ‘minimum rating constraint’ required to sustain a system and predict values for it based on various experimental rating datasets. Ultimately the approaches to user modeling illustrated in this and the following section are used to connect people. While a purely collaborative approach to recommendation is widely accepted and employed, it is riddled with endemic problems. User modeling must address more than just sparsity. For example, it is difficult to make connections to users with unusual or highly specific tastes. Furthermore, connecting users with similar interests who have rated different items (e.g., ‘we both read world politics online, but he ranked BBC.com webpages, while I ranked CNN.com pages’) is challenging. Over-specialization of evaluated artifacts, sometimes referred to as the ‘banana’ problem (Burke, 1999), arises since frequently purchased items, such as bananas in a grocery market basket, will always be recommended. Conversely, some products are seldomly bought more than a few times in a lifetime (e.g., automobiles) and thus suffer from a low number of evaluations. Over-specialization which is grounded in the exploration vs. exploitation dilemma can be addressed by occasionally forcing exploration. For instance, one can inject randomness (e.g., crossover and mutation in a genetic algorithm or epsilon in a reinforcement learning algorithm) into a model. Recommended artifacts also can be partitioned into hot and cold sets, where the latter is intended to foster exploration and increase the (rating) coverage of items in the system (Aggarwal et al., 1999). 3.1. Review of some representative projects The following collaborative-based systems employ many of the explicit user modeling techniques showcased above and illustrate what can be achieved with representations of users. People are connected in the following systems through statistical (Goldberg et al., 2000; Konstan et al., 1997), agent-oriented (Balabanovi´c and Shoham, 1997), and graphtheoretic (Aggarwal et al., 1999) approaches. GroupLens. GroupLens recommends Usenet news messages (Konstan et al., 1997). The system models users directly by explicitly eliciting and collecting ratings of messages through an independent newsreader. GroupLens is a project of the recommender systems research group at the University of Minnesota. Usenet news is a personal, voluminous, and ephemeral media (in comparison to movies) and thus an excellent candidate for collaborative filtering. A total of 250 people evaluated over 20,000 news articles (Resnick et al., 1994). GroupLens takes a statistical approach to making connections. The system predicts how a user seeking recommendation would rate an unrated article by computing a weighted average of the ratings of that message by users whose ratings were correlated with the user seeking recommendation. Correlation is computed with Pearson’s r coefficient. A research issue is deciding whether to provide personalized predictions (as GroupLens currently does) vs. personalized averages. Empirical research using Pearson’s r correlation coefficient revealed that ‘correlations between ratings and predictions is dramatically