Building similarity metrics reflecting utility in case-based reasoning

Fundamental to case-based reasoning is the idea that similar problems have similar solutions. The meaning of the concept of "similarity" can vary in different situations and remains an issue. Since we want to identify and retrieve truly useful or relevant cases for problem solving, the metrics of similarity must be defined suitably to reflect the utility of cases for solving a particular target problem. A framework for utility-oriented similarity modeling is developed in this paper. The main idea is to exploit a case library to obtain adequate samples of utility from pairs of cases. The task of similarity modeling then becomes the customization of the parameters in a similarity metric to minimize the discrepancy between the assessed similarity values and the utility scores desired. A new structure for similarity metrics is introduced which enables the encoding of single feature impacts and more competent approximation of case utility. Preliminary experimental results have shown that the proposed approach can be used for learning with a surprisingly small case base without the risk of over-fitting and that it yields stable system performance with variations in the threshold selected for case retrieval.