FCPFS: Fuzzy Granular Ball Clustering-Based Partial Multilabel Feature Selection With Fuzzy Mutual Information

Abstract

In the partial multilabel learning, incorrect labels are annotated because of their low quality and poor recognition. To decrease secondary errors in partial multilabel classification, this paper proposes a novel fuzzy granular ball clustering-based partial multilabel feature selection scheme with fuzzy mutual information. First, to overcome the defect that the traditional granular ball model cannot be applied to partial multilabel classification and its splitting rules are anomalous and stochastic, an objective function is designed by the fuzzy membership degree, the splitting rules and termination conditions are redesigned, and a new fuzzy granular ball clustering method using fuzzy k-means can be developed to preprocess partial multilabel data. Second, to reduce the impact of noise labels, the instance set of each granular ball is generated according to fuzzy granular ball clustering instead of neighborhood class, and the fuzzy similarity relationship between instances is constructed. Subsequently, granular ball-based fuzzy entropy measures and fuzzy mutual information and their properties are proposed in granular ball-based partial multilabel systems. Finally, the dependence and relevance between features and label sets are studied, the significance of features based on fuzzy mutual information is presented, and then a heuristic partial multilabel feature selection method is constructed to enhance the effect of partial multilabel data classification. Experiments on 18 partial multilabel datasets illustrate the availability of our method compared to other multilabel classification algorithms in its classification effect.