A composite entropy-based uncertainty measure guided attribute reduction for imbalanced mixed-type data.

In real-world scenarios, datasets generally exhibit containing mixed-type of attributes and imbalanced classes distribution, and the minority classes in the data are the primary research focus. Attribute reduction is a key step in the data preprocessing process, but traditional attribute reduction methods commonly overlook the significance of minority class samples, causing the critical information possessed in minority class samples to damage and decrease the performance of classification. In order to address this issue, we develop an attribute reduction algorithm based on a composite entropy-based uncertainty measure to handle imbalanced mixed-type data. To begin with, we design a novel oversampling method based on the three-way decisions boundary region to synthesize the samples of minority class, for the boundary region to contain more high-quality samples. Then, we propose an attribute measure to select candidate attributes, which considers the boundary entropy, degree of dependency and weight of classes. On this basis, a composite entropy-based uncertainty measure guided attribute reduction algorithm is developed to select the attribute subset for the imbalanced mixed-type data. Experimental on UCI imbalanced datasets, as well as the results indicate that the developed attribute reduction algorithm is significantly outperforms compared to other attribute reduction algorithms, especially in total AUC, F1-Score and G-Mean. [ABSTRACT FROM AUTHOR]