Abstract:

Self‑representation based unsupervised feature selection can handle unlabeled data without being affected by pseudo‑labeling. To ensure that such methods simultaneously achieve good robustness， preserve the local structure of samples， and select the most representative features， a new approach is proposed， and a corresponding iterative optimization algorithm is designed to compute its objective function. The method first identifies and processes outliers of samples， then combines the traditional self‑representation model with non‑convex sparse constraint and manifold regularization to form the target model， and puts the preprocessed data into the model for feature selection. Finally， the method uses the selected features for clustering. The proposed method is compared with seven methods on nine real data sets for experiments， and the experimental results show that the proposed method can effectively solve the unsupervised feature selection problem.

Key words: unsupervised feature selection, self?representation, robust, sparse, manifold regularization