Similarity-based inference as evidential reasoning

Abstract

The guiding principle underlying most approaches to similarity-based reasoning (SBR) is the common idea that “similar causes bring about similar effects”. We propose a probabilistic framework of SBR which is based on a formal model of this assumption. This model, called a similarity profile, provides a probabilistic characterization of the similarity relation between observed cases (instances). A probabilistic approach seems reasonable since it adequately captures the heuristic (and hence uncertain) nature of the above hypothesis. Taking the concept of a similarity profile as a point of departure, we develop an inference scheme in which instance-based evidence is represented in the form of belief functions. The combination of evidence derived from individual cases can then be considered as a problem of information fusion. In this connection, we also address the problem of rating individual cases, and of modulating their influence on the prediction which is finally derived.