The exponential growth and heterogeneity of multiview data have made dimensionality reduction indispensable. Unsupervised multiview feature selection (UMFS) has emerged as a pivotal solution. However, existing embedded UMFS methods still face several challenges, including poor pseudo-label quality, inadequate modeling of nonlinear structures, and heavy reliance on hyperparameters. To address these challenges, we propose a novel UMFS method named adaptive graph-guided feature decomposition (AGFD) for UMFS. It integrates feature selection into a nonnegative matrix factorization (NMF) framework and introduces a view-adaptive weighting mechanism. Different from the method of directly learning the view similarity matrix, AGFD constructs a global similarity matrix by refining predefined similarity matrices using Kullback-Leibler (KL) divergence, allowing it to better capture underlying nonlinear relationships across views. The adaptive design of AGFD reduces hyperparameter sensitivity and enhances interpretability. Extensive experiments on eight public multiview datasets demonstrate that AGFD outperforms existing methods, validating its effectiveness in UMFS tasks. The source code is available at: https://github.com/ahyuan/AGFD.
Yuan et al. (Thu,) studied this question.