State of Interest: Learning Personalized Utility Requirements

Defining similarity measures is a crucial task when developing CBR applications. Particularly, when employing utility-based similarity measures rather than pure distance-based measures one is confronted with a difficult knowledge engineering task. Especially if different users of a CBR system have different demands on the case retrieval this task becomes really complex and sophisticated. In such a scenario, identical queries may require different retrieval results depending on the context of the particular user because the utility of the cases for the current problem situation of the user may vary significantly. Consider a product recommendation system in e-Commerce. Here the users are customers with individual preferences with respect to the offered products. For example, some customers focus more on the price of a product while others are mainly interested in the technical properties. These preferences, for example, can be represented in form of attributes weights, i.e. they can be encoded into the similarity measure used to retrieve suitable products. However, this approach may significantly increase the knowledge engineering effort when developing a recommendation system based on CBR. Instead of defining one domain specific similarity measure, one has to define several measures that consider the specific preferences of individual customers or customer classes, respectively. But even if one is willed to put up with this additional effort, it is still an open question how to acquire the required knowledge. In our point of view, here a learning approach may help to facilitate both issues. Firstly, it may reduce the effort to define several similarity measures. Secondly, it is probably the only feasible way to obtain the required knowledge. To apply such an approach, one has to define one or several initial similarity measures that approximate the user specific utility measures as well as possible. During the use of the system one has to acquire feedback about the quality of the retrieval results from the users to learn more specific measures for each user or user class.