Learning Similarity Functions from Qualitative Feedback

The performance of a case-based reasoning system often depends on the suitability of an underlying similarity (distance) measure, and specifying such a measure by hand can be very difficult. In this paper, we therefore develop a machine learning approach to similarity assessment. More precisely, we propose a method that learns how to combine given local similarity measures into a global one. As training information, the method merely assumes qualitative feedback in the form of similarity comparisons, revealing which of two candidate cases is more similar to a reference case. Experimental results, focusing on the ranking performance of this approach, are very promising and show that good models can be obtained with a reasonable amount of training information.