Multi-view unsupervised feature selection with unified measurement of consistency and diversity

The low cost and high efficiency of multi-view unsupervised feature selection (MvUFS) have greatly stimulated research interest in this field. However, existing graph-based MvUFS methods typically focus solely on either the consistency or the diversity of multi-view data, let alone jointly considering both. Moreover, most approaches rely on matrix optimization while neglecting the exploration of higher-order correlations. In this work, we propose a novel multi-view learning framework for unsupervised feature selection to address these problems. First, a unified module is designed to jointly measure consistency and diversity, enabling the construction of a pure graph for each view. These pure graphs are then fused to generate a consensus graph, which, together with latent representations, mutually constrains and facilitates the learning of optimal pure graphs. Furthermore, we employ a low-rank tensor to preserve high-order correlations among views. The proposed methods are seamlessly integrated into a unified framework. Extensive experiments demonstrate that our model outperforms several state-of-the-art feature selection algorithms.