Unsupervised feature selection via anomaly-aware fuzzy graph fusion and diffusion multi-centroid learning

As an essential dimensionality reduction technique, unsupervised feature selection aims to learn a compact representation that preserves the intrinsic structure of high-dimensional unlabeled data. However, existing methods fail to characterize the fuzzy relationships among data points, and they typically rely on a single centroid to represent samples, which degrades the discriminative power of the learned representations. To address this limitation, a novel unsupervised feature selection method via anomaly-aware graph fusion and diffusion multi-centroid learning (AGFMCFS) is proposed. To enhance the reliability of the graph structure, an anomaly-aware mechanism based on the local outlier factor is introduced to quantify the credibility of each sample, thereby adaptively suppressing the interference of anomalous samples during graph construction. On this basis, multiple fuzzy graphs are integrated through a weighting mechanism, which contributes to a more faithful characterization of uncertain neighborhood relationships and further enhances the robustness of the learned manifold structure. Furthermore, a diffusion bipartite graph model is developed to capture the interactions between samples and multiple centroids in a propagative manner, yielding more discriminative feature representations. Extensive experiments on twelve benchmark datasets demonstrate that the proposed method outperforms a range of competitive approaches in clustering tasks, validating its effectiveness and robustness.