An Emerging Fuzzy Feature Selection Method Using Composite Entropy-Based Uncertainty Measure and Data Distribution

Feature selection based on neighborhood rough set is a noteworthy step in dealing with numerical data. Information entropy, proven in many theoretical analysis and practical applications, is a compelling feature evaluation method for uncertainty information measures. Nonetheless, information entropy replaces probability with uncertainty measure to evaluate the average amount of information and ignores the decision distribution of data, especially in describing the uncertainty in imbalanced data. This paper discusses an emerging method for the feature selection in fuzzy data with imbalanced data by presenting a local composite entropy based on a neighborhood rough set. Based on the neighborhood rough set model, we discuss a similar relation to describe the relationship between different objects in unbalanced fuzzy data. In this process, to fully consider the distribution characteristics of unbalanced data, we construct a local composite entropy for handling the fuzzy decision systems with uncertainty and decision distribution, which is proven to be monotonic. Moreover, to improve the selection efficiency, a local heuristic forward greedy selection algorithm based on the local composite measure is designed to select the optimal feature subset. Finally, experimental results on twelve public datasets demonstrate that our method has better classification performance than some state-of-the-art feature selection methods in fuzzy data.

Automated summary

This paper introduces a feature selection method based on neighborhood rough set and proposes a local composite entropy for dealing with imbalanced data in fuzzy data. Experimental results demonstrate that the proposed method achieves better classification performance in fuzzy data.