Unified Feature Selection Approach for Complex Data Based on Fuzzy β-Covering Reduction via Information Granulation.

Abstract

Feature selection, as an important step of data analysis, is widely used in the fields of data mining, machine learning, and artificial intelligence. It can not only effectively alleviate the curse of dimensionality and improve model performance but also enhance model interpretability. In the real world, data is usually complex such as different feature types, the presence of missing values, and so on. However, most existing feature selection approaches are only capable of handling data with a single feature type. To address the issue of feature selection under the environment of complex data, this article proposes a unified feature selection (UFS) approach for complex data based on fuzzy $\beta $ -covering reduction via information granulation. To begin with, several monotonic uncertainty measures for fuzzy $\beta $ -covering are constructed from the viewpoints of algebra and information theory. Based on the proposed measures, two forward heuristic algorithms are designed for fuzzy $\beta $ -covering reduction. Meanwhile, the complex data with multiple features is represented by fuzzy $\beta $ -covering via information granulation. On this basis, a UFS approach is put forward for complex data. Finally, the effectiveness and superiority of the proposed approach are verified through a series of experiments compared with 12 state-of-the-art feature selection approaches.