A Design of Case-Based Dynamic Feature Weighting Method using Local and Global Selection

摘要

Case-based learning algorithms are lazy machine learning methods that store problem solving experiences. When given a new problem, they retrieve cases whose problems were similar, and reuse their classifications. Lazy learning methods have several advantages in comparison to eager approaches, but also have some drawbacks. For example, case retrieval is sensitive to irrelevant features, which can degrade classification accuracy. Although feature weighting methods have been proposed to combat this problem, most of these use global weighting schemes. We introduce a new wrapper weighting method, named CaDFeW (CAse-based Dynamic FEature Weighting), which instead uses locally varying weight vectors. It tracks the classification performance of randomly generated feature weight vectors and dynamically constructs a weight vector for a given problem. Its advantages are (1) lower processing costs than other wrapper methods and (2) a flexible architecture. We evaluate CaDFeW's accuracy, and its integration with a global weighting method, on several classification tasks, and introduce a new definition of input dependency to explain its results. We found that the relative performance of a local feature weighting method versus a global feature weighting method is an increasing function of the domain's input dependency.