Class-specific semi-supervised feature selection with fuzzy convex balling information granularity

Hengrong Ju, Jiarong Li, Weiping Ding, Xiaoxue Fan, Jiashuang Huang, Suping Xu, Xibei Yang

Abstract: Feature selection is widely used method for dimension reduction in data mining. Traditional granular computing-based feature selection methods typically rely on fully labeled data. However, in real world big data mining, only partial data contains label information. In order to solve this issue, this paper proposes a novel class-specific semi-supervised feature selection approach with fuzzy convex balling information granulation. First, in light of Pythagorean theorem, two adaptive neighborhood spaces are identified, followed by constructing fuzzy convex balling information granules using the shared neighborhood space. Second, a novel class-specific based fuzzy decision learning approach is studied, in which a fuzzy convex balling information granule is created for each class. Third, we develop fuzzy convex balling information granule-based measures to evaluate the performance of features. Finally, the proposed class-specific semi-supervised feature selection approach is validated on a number of public datasets from three perspectives, involving feature quality assessment, fuzzy decision accuracy, and brain region selection for schizophrenia. The experimental results show that the feature selection method proposed in this paper outperforms existing algorithms in processing semi-supervised data.