Weighting Features

Many case-based reasoning algorithms employ derivatives of the k-nearest neighbor (k-NN) classiier for case retrieval. Several studies have shown that its similarity function is sensitive to imperfect feature sets (i.e., containing irrelevant, redundant, interacting, or noisy features). Many proposed methods attempt to reduce this sensitivity by parameterizing k-NN's similarity function with feature weights. We focus on the subset of these methods that automatically assign weight settings using little or no domain-speciic knowledge. Our goal is to understand the relative capabilities of these methods for given sets of conditions. We rst describe a ve-dimensional framework that categorizes automated weight-setting methods. Next, we empirically compare methods along one of these dimensions. Finally, we summarize our results with four hypotheses and describe additional empirical evidence to support them. Our investigation revealed that most methods correctly assign low weights to completely irrelevant features, and that methods which use performance feedback to assign weight settings demonstrate three advantages over other methods (i.e., they require less pre-processing, perform better in the presence of interacting attributes, and support faster learning rates).