关键词: 模糊等价关系, 联合互信息, 最大最小准则, 特征选择

Abstract: Feature selection is the method that selects feature subset that has strong relevancy between features and classification and smallest redundancy among features from feature set. This can improve the classifier’s computational efficiency, and enhance the classifier’s generalization, and therefore increase classification accuracy. However, the relevance and redundancy evaluation criteria based on mutual information has the following problems in the practical applications：（1） It is difficult to calculate the probability of a variable and the feature’s information entropy; （2） The approach based on mutual information tends to choose features which have more values; （3） The method measuring redundancy between candidate features and selected feature subset based on cumulative addition with higher dimension data sets always is invalid. To solve the above problems, the feature evaluation criteria based on Normalized Fuzzy Joint Mutual Information Maximum（NFJMIM） is proposed in this paper. Firstly, the entropy, conditional entropy, joint entropy of a variable are calculated based on fuzzy equivalence relation. Secondly, the feature’s importance is evaluated base on NFJMIM. Finally, using the established criteria, forward greedy search approach is used for searching feature subset. Several experiments using UCI machine learning repository prove that the proposed algorithm can effectively select effective feature subset, and can significantly improve the classification accuracy.

Key words: fuzzy equivalence relations, joint mutual information, the maximum and minimum criteria, feature selection