An improved fuzzy mutual information feature selection for classification systems

摘要

Classification systems are sensitive to input data, especially for datasets with a lot of undesirable features. Selecting relevant features and avoiding irrelevant or redundant features builds effective systems. Fuzzy Mutual Information measures the relevance and redundancy of features. Although it can deal directly with continuous data without discretization, it still requires more computation and storage space. In this paper, we propose an improved fuzzy mutual information to solve this problem. Furthermore, we integrate it with normalized max-relevance and min-redundancy (mRMR) approach. It does not only select the relevant features but also avoids the redundancies with respect to the domination between them. Our experiment was evaluated according to storage, stability, classification accuracy, and the number of selected features. Based on 12 benchmark datasets, experimental results confirm that our proposed method achieved better results.