Effect of data weighting methods on the performance of fuzzy classification systems

In this paper, we examine the performance of fuzzy rule-based systems for pattern classification problems. We assume that each training pattern has a weight that corresponds to the importance of the pattern. A fuzzy classification system is constructed by generating fuzzy if-then rules from the weighted training patterns. We consider three weighting methods: we first consider a random weighting method that assigns a random value to each of training patterns. Next a class-based weighting method is considered where weights are determined depending on the class of training patterns. The third one is an overlap-based weighting method where weights reflect the degree of overlap between different classes. We use several real-world data sets as classification problems in the computer simulations in this paper. In the construction of fuzzy classification systems, we use two fuzzy rule-generation methods. One method determines the consequent class of fuzzy if-then rules only from the class information of compatible patterns. In the other method, weights of compatible patterns to fuzzy if-then rules are also used together with the class information. We show the advantages and disadvantages of the three weighting methods. The effect of the weighting methods on the generalization ability of fuzzy classifications is also presented.