Leveraging Transformer-based autoencoders for low-rank multi-view subspace clustering

Deep multi-view subspace clustering is a hot research topic, aiming to integrate multi-view information to produce accurate cluster prediction. Limited by the inherent heterogeneity of distinct views, existing works primarily rely on view-specific encoding structures for representation learning. Although effective to some extent, this approach may hinder the full exploitation of view information and increase the complexity of model training. To this end, this paper proposes a novel low-rank multi-view subspace clustering method, TALMSC, backed by Transformer-based autoencoders. Specifically, we extend the self-attention mechanism to multi-view clustering settings, developing multiple Transformer-based autoencoders that allow for modality- agnostic representation learning. Based on extracted latent representations, we deploy a sample-wise weighted fusion module that incorporates contrastive learning and orthogonal operators to formulate both consistency and diversity, consequently generating a comprehensive joint representation. Moreover, TALMSC involves a highly flexible low-rank regularizer under the weighted Schatten p-norm to constrain self-expression and better explore the low-rank structure. Extensive experiments on five multi-view datasets show that our method enjoys superior clustering performance over state-of-the-art methods.