Fuzzy Rule-Based Approaches to Dimensionality Reduction

Abstract

In this talk we deal with the problem of dimensionality reduction in a fuzzy rule-based framework. We consider dimensionality reduction through feature extraction as well as through feature selection. For the former approach, we use Sammon's stress function as a criterion for structure-preserving dimensionality reduction. For feature selection we propose an integrated framework, which embeds the feature selection task into the classifier design task. This method uses a novel concept of feature modulating gate and it can exploit the subtle nonlinear interaction between the tool (here a fuzzy rule based system), the features and the task at hand. This method is then extended to Takagi-Sugeno (TS) model for function approximation/prediction problem. The effectiveness of these methods is demonstrated using several data sets.