Acquiring similarity cases for classification problems

Abstract

The situation assessment and similarity components of the interpretive case-based reasoning process are integral for a successful case retrieval. However, for classification problems there are domains where it can be difficult to define sets of relevant features to extract from a problem description. Likewise it is not always obvious which of these features to apply to the similarity assessment process and what, if any, weights they should be given. We suggest learning the concept of similarity by training on a set of past situations. Rather then develop a general function, we store the knowledge gained in individual similarity comparisons as similarity cases. These similarity cases define a similarity space that can be searched to identify how new problem situations can be classified. This paper describes our approach of acquiring similarity cases in the context of a straightforward classification task. A proof of concept system was built that creates similarity cases from a repository of known spam email messages and can use the similarity cases to classify unknown messages as positive or negative examples of spam. © Springer-Verlag Berlin Heidelberg 2005.