Generating Fuzzy Rules from Weighted Training Patterns for Pattern Classification

Abstract

This paper proposes a fuzzy rule-generation method from weighted training patterns. We assume that each training pattern has a weight that corresponds to its importance. The weight is treated as a cost of a misclassified/rejected pattern. The weights of the training patterns are used in the generation process of fuzzy if-then rules. We formulate the problem of constructing classification system as minimization of a total cost and an error rate. The proposed method handles training patterns with large weights more importantly than those with small weights. Our objective is to construct a fuzzy classification system that decreases a total cost and error rate. In the classification of an unseen pattern, we take the maximum value of the product of the compatibility and the grade of certainty. In computational experiments, we compare the conventional construction method of fuzzy classification system with the proposed method. We show that a total cost and an error rate are reduced by the proposed method. The performance of the proposed method is also compared with that of nearest neighbor classifiers.