Semi-supervised feature selection via fuzzy C-means clustering and simulated annealing optimization

Abstract Feature selection is a critical step in machine learning, as it helps minimize redundant features, thereby enhancing model performance and interpretability. In practice, partially labeled datasets are common. When some data labels are unavailable, semi-supervised feature selection has become the predominant approach. Existing methods for semi-supervised feature selection face two major challenges when dealing with partially labeled data: (1) Defects in pseudo-label generation quality: Traditional hard clustering methods, such as K-means, generate pseudo-labels that are susceptible to noise interference; (2) Inefficient feature subset optimization: Conventional optimization strategies are prone to becoming trapped in local optima within high-dimensional spaces and often lack deep integration with semi-supervised evaluation criteria, resulting in low search efficiency and slow convergence. To address these issues, this paper proposes a novel semi-supervised feature selection algorithm named Semi-FCMSA, which integrates fuzzy clustering with global optimization. First, labeled samples are utilized to initialize a Fuzzy C-Means (FCM) clustering algorithm, generating soft membership degrees as pseudo-labels for unlabeled samples. Second, we introduce a triple-criterion metric named REL-RED-IR to quantify feature relevance, redundancy, and irrelevance. Finally, simulated annealing global search is combined with local feature permutation operations to optimize the feature subset. Experiments conducted on multiple UCI datasets demonstrate that the proposed Semi-FCMSA algorithm achieves superior classification performance compared to four benchmark methods.