A Study on Weighting Training Patterns for Fuzzy Rule-Based Classification Systems

In this paper, we examine the effect of weighting training patterns on the performance of fuzzy rule-based classification systems. A weight is assigned to each given pattern based on the class distribution of its neighboring given patterns. The values of weights are determined proportionally by the number of neighboring patterns from the same class. Large values are assigned to given patterns with many patterns from the same class. Patterns with small weights are not considered in the generation of fuzzy rule-based classification systems. That is, fuzzy if-then rules are generated from only patterns with large weights. These procedures can be viewed as preprocessing in pattern classification. The effect of weighting is examined for an artificial data set and several realworld data sets.