Multi-view Unsupervised Feature Selection Guided by Latent Representation and Tensor Learning

Multi-view unsupervised feature selection refers to the process of selecting the most representative or relevant feature subset from a dataset containing multiple views. Most existing methods in this field use the original features to construct similarity graphs, but real datasets often contain noise and redundant information. The method of constructing adaptive graphs and spectral embedding both obtain the local structure of data and ignore the global information. On the other hand, they usually obtain consistency information through multi-graph fusion strategies or the construction of consistency graphs. However, due to the heterogeneity of view features, such methods are prone to lose original information and it is difficult to identify shared similar structures. To address these problems, we propose a model named Multi-view Unsupervised Feature Selection Guided by Latent Representation and Tensor Learning (TLMvFS). Our approach maps raw data into a latent space to capture semantic structural information. Subspace learning is then applied to the novel latent representation matrix to eliminate irrelevant features, enabling a more accurate capture of the intrinsic data structure. Currently, spectral representation is introduced to preserve shared local structures. To model correlations across multiple views, we adopt low-rank tensor learning to characterize high-order correlations. Furthermore, a \(\ell_{2,p}\)-norm regularization is incorporated into the feature regression framework to enhance robustness. Extensive experiments conducted on six benchmark datasets demonstrate that our method outperforms multiple state-of-the-art models.