be ranked by attempting to maximise agreements with this learned preference function see also the work of [13 Finally, it is worth highlighting recent similarity learning work by O'Sullivan et al [66, 67, 68]. This work has not been used directly by case-based recommenders but in stead has been used to improve the quality of collaborative filtering recommenders(see Chapter 9[83] in this book) by using case-based style similarity metrics when evalu ating profile similarities. Normally a collaborative filtering recommender system can only evaluate the similarity between two profiles if they share ratings. For example in a TV recommender two users that have both rated ER and Frasier can be compared. But if one user has only rated ER and the other has only rated Frasier then they cannot be ompared. O Sullivan et al. point out that the ratings patterns within a collaborative fil tering database can be analysed to estimate the similarity between programmes like ER and Frasier. They show that by using data-mining techniques it is possible to discover hat, for example, 60% of the people who have liked Er have also liked Fraiser, and use this as a proxy for the similarity between these two programmes. They demonstrate how significant improvements in recommendation accuracy can be obtained by using these similarity estimates with more sophisticated case-based profile similarity metrics 11.2.4 Single-Shot Recommendation Many case-based recommenders operate in a reactive and single-shot fashion, present ng users with a single set of recommendations based on some initial query; thus the user is engaged in a single(short-lived) interaction with the system. For example, the Analog Devices OpAmp recommender presents a user with a set of available Op Amps that closely match the user's query [100, 103]. The online property recommender re- ferred to earlier operate similarly, responding with a selection of suitable apartments in response to a user's rental constraints; see also the DubLet system by [40] The point to make here is that single-shot recommendation has its shortcom- ings. In particular, if users do not find what they are looking for among the initial recommendations--as is frequently the case--then their only option is to revise their query and start again. Indeed the pure similarity-based nature of most case-based rec- ommender systems increases the chances of this happening in certain situations be cause, as we discussed earlier, the top ranked recommendations may differ from the target query in more or less the same ways. As a result they will be very similar to each other---they will lack diversity-and if the user doesn't like the first recommenda tion she is unlikely to be satisfied with the similar alternatives either. In the remaining sections of this chapter we will explore how this simple model of case-based recommen dation has been extended to provide a more sophisticated recommendation framework, one that provides for more sophisticated interaction between recommender and user, generating personalized recommendations that are more diverse, through an extended dialog with the user
352 B. Smyth be ranked by attempting to maximise agreements with this learned preference function; see also the work of [13]. Finally, it is worth highlighting recent similarity learning work by O’Sullivan et al. [66, 67, 68]. This work has not been used directly by case-based recommenders but instead has been used to improve the quality of collaborative filtering recommenders (see Chapter 9 [83] in this book) by using case-based style similarity metrics when evaluating profile similarities. Normally a collaborative filtering recommender system can only evaluate the similarity between two profiles if they share ratings. For example in a TV recommender two users that have both rated ER and Frasier can be compared. But if one user has only rated ER and the other has only rated Frasier then they cannot be compared. O’Sullivan et al. point out that the ratings patterns within a collaborative filtering database can be analysed to estimate the similarity between programmes like ER and Frasier. They show that by using data-mining techniques it is possible to discover that, for example, 60% of the people who have liked ER have also liked Fraiser, and use this as a proxy for the similarity between these two programmes. They demonstrate how significant improvements in recommendation accuracy can be obtained by using these similarity estimates with more sophisticated case-based profile similarity metrics. 11.2.4 Single-Shot Recommendation Many case-based recommenders operate in a reactive and single-shot fashion, presenting users with a single set of recommendations based on some initial query; thus the user is engaged in a single (short-lived) interaction with the system. For example, the Analog Devices OpAmp recommender presents a user with a set of available OpAmps that closely match the user’s query [100, 103]. The online property recommender referred to earlier operate similarly, responding with a selection of suitable apartments in response to a user’s rental constraints; see also the DubLet system by [40]. The point to make here is that single-shot recommendation has its shortcomings. In particular, if users do not find what they are looking for among the initial recommendations—as is frequently the case—then their only option is to revise their query and start again. Indeed the pure similarity-based nature of most case-based recommender systems increases the chances of this happening in certain situations because, as we discussed earlier, the top ranked recommendations may differ from the target query in more or less the same ways. As a result they will be very similar to each other—they will lack diversity—and if the user doesn’t like the first recommendation she is unlikely to be satisfied with the similar alternatives either. In the remaining sections of this chapter we will explore how this simple model of case-based recommendation has been extended to provide a more sophisticated recommendation framework, one that provides for more sophisticated interaction between recommender and user, generating personalized recommendations that are more diverse, through an extended dialog with the user
11 Case-Based Recommendation 353 11.3 Similarity and beyond Let us look at a concrete example of the diversity problem referred to above. Consider a vacation recommender where a user submits a query for a 2-week vacation for two in the sun, costing less than $750, within 3 hours flying time of Ireland, and with good night-life and recreation facilities on-site. The top recommendation returned is for an apartment in the Hercules complex in the Costa Del Sol, Spain, for the first two weeks in July. A good recommendation by all accounts, but what if the second, third, and fourth recommendations are from the same apartment block, albeit perhaps for differ ent two-week periods during the summer, or perhaps for different styles of apartments? While the k(k=4 in this case) best recommendations are all very similar to the target query, they are also very similar to each other. The user has not received a useful set of Iternatives if the first recommendation is unsuitable. This scenario is not uncommon ○○ o0oo○○ 0○○ ○○ a)○ (b)○○ Fig. 11.7. Similarity vs diversity during case retrieval: (a) a case base with highlighted target query, t;(b)a conventional similarity-based retrieval strategy returns the cases that are individu lly closest to the target query, thus limiting their potential diversity; (c)an alternative retrieval strategy that balances similarity to the target and the relative diversity of the selected cases pro- duces a more diverse set of recommendations in recommender systems that employ similarity-based retrieval strategies: they often produce recommendation sets that lack diversity and thus limit user options(see Figure 11.7(a&b). These observations have led a number of researchers to explore alternatives to similarity-based retrieval, alternatives that attempt to explicitly improve recommen dation diversity while at the same time maintaining query similarity 11.3.1 Similarity vS Diversity How then can we improve the diversity of a set of recommended cases, especially since many of the more obvious approaches are likely to reduce the similarity of the selected I Incidentally, related concerns regarding the primacy of similarity in other forms of case-based reasoning have also come to light, inspiring many researchers to look for alternative ways to judge the utility of a case in a given problem solving context(e.g [7, 19, 34, 45, 49, 91)). For example, researchers have looked at the importance of adaptability alongside similarity, arguing that while a case may appear to be similar to a target problem, this does not mean it can be successfully adapted for this target(see[49, 91D
11 Case-Based Recommendation 353 11.3 Similarity and Beyond Let us look at a concrete example of the diversity problem referred to above. Consider a vacation recommender where a user submits a query for a 2-week vacation for two in the sun, costing less than $750, within 3 hours flying time of Ireland, and with good night-life and recreation facilities on-site. The top recommendation returned is for an apartment in the Hercules complex in the Costa Del Sol, Spain, for the first two weeks in July. A good recommendation by all accounts, but what if the second, third, and fourth recommendations are from the same apartment block, albeit perhaps for different two-week periods during the summer, or perhaps for different styles of apartments? While the k (k = 4 in this case) best recommendations are all very similar to the target query, they are also very similar to each other. The user has not received a useful set of alternatives if the first recommendation is unsuitable. This scenario is not uncommon c3 c2 c1 t c3 c2 c1 t (a) (b) (c) t Fig. 11.7. Similarity vs diversity during case retrieval: (a) a case base with highlighted target query, t; (b) a conventional similarity-based retrieval strategy returns the cases that are individually closest to the target query, thus limiting their potential diversity; (c) an alternative retrieval strategy that balances similarity to the target and the relative diversity of the selected cases produces a more diverse set of recommendations. in recommender systems that employ similarity-based retrieval strategies: they often produce recommendation sets that lack diversity and thus limit user options (see Figure 11.7(a&b)). These observations have led a number of researchers to explore alternatives to similarity-based retrieval, alternatives that attempt to explicitly improve recommendation diversity while at the same time maintaining query similarity. 1 11.3.1 Similarity vs. Diversity How then can we improve the diversity of a set of recommended cases, especially since many of the more obvious approaches are likely to reduce the similarity of the selected 1 Incidentally, related concerns regarding the primacy of similarity in other forms of case-based reasoning have also come to light, inspiring many researchers to look for alternative ways to judge the utility of a case in a given problem solving context (e.g. [7, 19, 34, 45, 49, 91]). For example, researchers have looked at the importance of adaptability alongside similarity, arguing that while a case may appear to be similar to a target problem, this does not mean it can be successfully adapted for this target (see [49, 91])