Data-based Fuzzy Rules Classification Algorithm

Classification and feature selection are hot topics in pattern recognition and machine learning. Classification is a technique that modeled by labeled data and then applies the model to identify the label of unlabeled data. Feature selection, as the pre-processing part of the classification technique, can remove redundant features and noisy features, simplify the building process of a classifier, which is becoming more and more important in the classification technique.A general algorithm was proposed by Lixin Wang and Jerry M. Mendel in 1992 to generate fuzzy rules from numerical data. In this paper, we design a new data-based fuzzy classifier based on the classic algorithm. In the pre-processing phase of the classifier, we propose a neighborhood-based greedy feature selection algorithm to select the most important r features, which improve the building efficiency of the classifier. In the designing phase of the classifier, we propose a new way to compute the confidence degree, counting the number of training samples which the identical rule is extracted from. We have compared the new method with traditional computing method of confidence degree in 10 UCI data sets by Friedman test and Holm test, and the results of the statistical tests supporting comparative analysis shows that the new method performs significantly better than the traditional method. Then the rule-base is pruned to delete the rules which cause bad performance.We have experimented with 7 UCI data sets to test the performance of the proposed classifier and have compared the results with the other five decision tree classifiers. It has been showed that the proposed method represents a highest accuracy and a comparable number of rules than those being obtained by other five methods. The rule-base extracted by the new algorithm is much simpler, easier to understand and easier to interpretable.