Incomplete multi-view clustering via self-attention networks and feature reconstruction

Over the past few years, numerous deep learning-based methods have been proposed for incomplete multi-view clustering. However, these approaches overlook two crucial issues. First, they focus solely on the global information contained in the latent representations derived from deep networks, neglecting the importance of local focal points. Second, while leveraging consistent or complementary inter-view information for cross-view learning, they disregard the intrinsic relationships among different samples within the same view. To address these concerns, this manuscript presents an original approach: incomplete multi-view clustering based on self-attention networks and feature reconstruction (SNFR). Specifically, SNFR initially employs self-attention networks to emphasize the pivotal information within views, aiming to reduce the inter-view reconstruction loss. Subsequently, an improved entropy weighting method is applied to reconstruct the feature relationships among the diverse samples within the same view, thereby facilitating consistent cross-view information learning. Our proposed method is evaluated on six widely used multi-view datasets through extensive experiments, highlighting its remarkable superiority over the alternative approaches in terms of clustering performance