Using a Relevance Model for performing Feature Weighting

Feature Weighting is one of the most difficult tasks when developing Case Based Reasoning applications. This complexity grows when dealing with ill-defined wide domains with a sparse case base. Moreover, most widely-used feature selection and feature weighting methods assume that features are either relevant in the whole instance space or irrelevant through-out. However, it is often the case that specific features are only relevant within the context of other features’ values (i.e., feature is relevant if feature , but irrelevant if ). Therefore, features’ weight and relevance are Context-Sensitive. This paper defines a model to capture complex relations of relevance among features using domain knowledge, and suggests a method for mapping the knowledge captured in the model into feature weights, performing in this way Context-Sensitive Feature Weighting. The method we suggest is suitable for those domains where no instance based learning algorithm or observation based approach is applicable.