Unsupervised feature selection based on adaptive latent representation learning and multi-group data similarity

Latent representation learning has become a common method in feature selection. Most algorithms typically learn a fixed low-dimensional pseudo-label matrix from the traditional similarity matrix, which makes it challenging to learn accurate latent representations. To solve this problem, an unsupervised feature selection method based on adaptive latent representation learning and multi-group data similarity (AMUFS) is proposed. Firstly, AMUFS introduces adaptive latent representation learning, and adaptively adjusts the scale of the latent representation matrix through sparse constraints. It can better capture the latent representation of the data, avoiding the issue where the fixed latent factor is not flexible enough. Secondly, the multi-group data similarity based on pre-grouping is designed. The pre-grouping clustering algorithm is used to cluster the original data into several groups, and multi-group data similarity is constructed based on the pre-grouping results. This approach reduces noise and better reflects the correlation information between samples. Finally, the ℓ_2,1 loss is used to constrain the sparse regression model, and the non-convex constraint is applied to sparse regularization. This improves accuracy and feature robustness enabling AMUFS to select more discriminative solutions. AMUFS is evaluated against six distinct feature selection algorithms using seven publicly accessible datasets. The experimental results demonstrate that AMUFS achieves superior clustering accuracy and normalized mutual information compared to the other methods.