From: FLAIRS-02 Proceedings. Copyright © 2002, AAAI (www.aaai.org). All rights reserved. Using Collaborative Filtering Data in Case-Based Recommendation Derry O’ Sullivan and David Wilson and Barry Smyth Smart Media Institute University College Dublin fdermot.osullivan,david.wilson,[email protected] Abstract cases. However, this presents a number of challenges that are side-effects of the so-called sparsity problem In the context of PTV, an applied recommender system operating in the TV listings domain, we are examining that plagues many collaborative filtering systems. The the potential benefits in merging case-based and collab- sparsity problem tells us that, on average, two users are orative filtering (CF) recommendation techniques by unlikely to have rated many of the same items and so developing case-based reasoning (CBR) methods that there will be little direct overlap between their profiles. employ collaborative filtering style ratings profiles di- This is problematic when it comes to case retrieval, be- rectly as cases. Doing so presents a number of chal- cause it means that there is no direct way to measure lenges, both in applying a case-based perspective to the similarity between two profile cases unless we have collaborative filtering, and in addressing the sparsity access to similarity knowledge that allows us to compare problem that plagues many collaborative filtering sys- non-identical profile items. Unfortunately this similar- tems. This paper expands on earlier CBR views of ity knowledge is usually hard to acquire. collaborative filtering, identifies problems and opportu- nities for similarity maintenance therein, and proposes Moreover, even when we do have access to similarity and evaluates methods for mining and applying new knowledge it is not clear how it should be used to deter- similarity knowledge. mine case similarity. Ordinarily, in CBR systems, the case retrieval step is facilitated by the so-called align- ment assumption which states that two cases can be Introduction compared on a feature by feature basis. But when com- In the context of recommender systems, case-based and paring profile cases it may be possible to relate individ- collaborative filtering techniques have been, at the same ual features (rated items) in the candidate case to many time, viewed as both complimentary and contrasting. features in the target case; in other words there may be They can both be seen as lazy similarity-based reason- no clear one-to-one correspondence between items. ing techniques. Case-based methods generate recom- In this paper we address both of these challenges by mendations by prioritizing cases that are similar to ones describing how similarity knowledge can be mined from that the target user has preferred in the past, while col- the profile cases themselves and by advancing a novel laborative filtering methods prioritize cases that have similarity metric for comparing non-aligned cases. We been liked by users similar to the target user. How- demonstrate the effectiveness of these ideas through an ever these techniques differ significantly in the way that experimental evaluation and argue that our technique similarity is assessed and the type of data on which the has broader implications for the CBR community as a similarity computation is based. Case-based methods whole. rely on rich feature-based representations and sophisti- cated similarity metrics that make use of heterogeneous Setting the Scene similarity measures in order to deal with the various PTV is a deployed recommender system operating features that can make up a case. Collaborative fil- in the TV listings domain that combines case-based tering methods, in contrast, make use of very simple and collaborative filtering techniques by interleaving correlation-based similarity techniques to measure the the recommendations (TV programmes) from each ap- similarity between user profiles that are typically little proach to produce personalized TV guides for each indi- more than ratings lists; that is, content item identifiers vidual user (Smyth & Cotter 2001). The key to PTV’s plus an associated preference rating. personalization facility is an accurate database of inter- In this paper we are interested in developing a case- actively acquired user preference profiles that contain based recommendation technique that uses ratings- collaborative filtering style ratings lists. These are em- based profiles (a la collaborative filtering) directly as ployed directly in the collaborative filtering component Copyright °c 2002, American Association for Artificial In- and by transformation to a content summary profile telligence (www.aaai.org). All rights reserved. schema for matching in the case-based component. We FLAIRS 2002 121 would like to improve recommendations in the overall The Similarity Coverage Problem system by taking this combination a step further with Taking a CBR view of CF, we must begin to address a case-based view of the collaborative component it- weaknesses in the underlying technique, in particular self. Within this view, we have identified opportunities the sparsity problem. In collaborative filtering, the spar- for maintaining and improving collaborative recommen- sity problem tells us that on average two users are un- dations and for further developing the relationship be- likely to have rated many of the same items and so there tween CBR and CF. We are mindful, however, that the will be little direct overlap between their profiles. From collaborative filtering component is desirable, in part, a CBR perspective, there may be no direct way to mea- because it can provide diverse and high-quality recom- sure the similarity between two such profile cases. The mendations with minimal knowledge-engineering effort. tacit assumption in CBR has generally been that the One of the goals of this work is to strike a good balance similarity metric will cover all potential retrieval situa- between improving the collaborative recommendations tions. It may do so poorly, necessitating for example, and the amount of development overhead in doing so. learning appropriate feature weights, but it would pro- vide complete coverage. This is not the situation with A CBR Perspective on CF profile cases, some of which may not be comparable at all, and it presents us with a similarity coverage prob- Recent CBR research has started to investigate the rela- lem, akin to the case-base coverage problem (Smyth & tionship between CBR and CF (e.g., (Hayes, Cunning- McKenna 1998) that has received a great deal of recent ham, & Smyth 2001; Burke 2000)). Hayes et al. (2001) attention. forward a view of Automated Collaborative Filtering Addressing the similarity coverage problem is a main- as a lazy CBR case-completion process. We adopt this tenance issue for CBR systems, directed at the simi- view here, emphasizing two important differences from larity knowledge container (Wilson & Leake 2001). In typical CBR practice. order to maximize similarity coverage, we need to em- First, there is no case solution distinct from case ploy techniques to derive and extend similarity knowl- specification. This follows research developments in edge. In our case, we need to gain access to similarity CBR from dialog-based/conversational systems (e.g., knowledge that allows us to compare non-identical pro- (Doyle & Cunningham 2000; Aha, Breslow, & Mu˜noz- file items. Avila 2001)) to case/information-completion systems (e.g., (Burkhard 1998; Lenz, Auriol, & Manago 1998)). Mining Similarity Knowledge In dialog-based systems, system-guided user interac- tion is used to fill out the problem specification dur- There are many automated techniques that could be ing problem-solving, but the goal remains to find a dis- used to derive various sorts of similarity knowledge. tinct solution based on the problem specification. Case- The initial approach we have chosen is to apply data completion systems, on the other hand, are dialog-based mining techniques (see (Hipp & Nakhaeizadeh 2000) systems in which the elicitation of problem features is for an overview)), in particular the Apriori algorithm both the means and the end. For example, incremen- (Agrawal et al. 1995), to extract association rules tally filling in the next aspect of a design may be sup- between programmes in PTV user-profile cases. By ported by reasoning from the current partial context discovering relationships between programmes beyond of the design so far (Leake & Wilson 2001). Cases, simple direct overlap, we may be able both to cover then, are particular points in an incremental develop- more potential profile matches and to make more in- ment process, and the case-based reasoning cycle is ap- formed recommendations. For example, a person that plied repeatedly to support their refinement, which is likes X-Files and Frasier would not normally be compa- very much in the spirit of case life-cycle models (Minor rable to a person that likes Friends and ER, but mining & Hanft 2000). We also note that some case life-cycles a relationship between Frasier and Friends would pro- may be expected to end, as in completing a design, vide a basis for profile matching. while other life-cycles may be expected to persist indef- The association rules are of the form A ) B, where initely, as in recommendation. A and B are sets of items (television programmes). In Second, ratings-based profiles cases do not conform data mining terms, whenever a transaction (case) T to the typical structure of parallel features with corre- contains a certain itemset (set of programmes) A, then sponding heterogeneous similarity measures. This is a the transaction probably contains another itemset B. direct result of the sparsity typical in collaborative sim- The probability that a given rule holds, rule confidence, ilarity spaces, and we discuss the ramifications in the is the percentage of transactions containing B given following section. Thus the alignment assumption that that A occurs: two cases can be compared on a feature by feature basis P (B ⊆ T jA ⊆ T ) (1) does not necessarily hold in such systems.