Multi-view Unsupervised Feature Selection With Joint Multi-subspace Robust Learning

Multi-view unsupervised feature selection aims to extract representative and discriminative features from multiple views to improve model performance on complex datasets. However, existing methods have the following issues: The affinity graph constructed with the original data as input reduces the flexibility of feature selection. They focus on mining the intrinsic structural information of multi-view data, adopt fusion strategies to mine consistent information, but ignore the exploration of high-order correlations. The original data space cannot effectively represent the intrinsic geometric structure of the data. To address these problems, we propose a new model for multi-view unsupervised feature selection, called Multi-view Unsupervised Feature Selection with Joint Multi-Subspace Robust Learning (MSRMFS). This method mines the underlying structure of the data by fusing multiple subspace representations. Specifically, it combines subspace representation, subspace representation of latent consistency, and augmented view subspace representation, and stacks each self-representation matrix into a three-dimensional tensor. We employ a learning mechanism to adaptively capture the shared structure and obtain consistent clustering indicator structures through the spectral embedding of weighted Laplacian graphs across different views, which is then incorporated into a sparse feature selection model. Additionally, to make the subspace representations more robust, we constrain the reconstruction error with the l2,1 norm. Extensive experiments on five datasets demonstrate that our proposed approach outperforms several state-of-the-art models.