Abstract: Most HSIC-based feature selection methods are subject to the following limitations. First, these methods are typically only applied to labeled data, which is not feasible since most of the data in real-world applications is unlabeled. Second, existing HSIC-based unsupervised feature selection methods only address the general correlation between the selected features and the output values representing the underlying clustering structure, while ignoring the redundancy between different features. To address these issues, a new unsupervised feature selection method based on HSIC(UFSHSIC)is proposed, which utilizes HSIC as a correlation criterion to explore the correlation between features and the overall sample structure, as well as the redundancy between features. Experimental comparison with other classical feature selection methods on multiple real datasets shows that the proposed method can effectively perform feature selection from unlabeled samples, and the selected feature subset produces equivalent or better performance than supervised feature selection methods.

Key words: unsupervised feature selection, Hilbert-Schmidt independence criterion(HSIC), kernel method, machine learning, feature redundancy