Multi-View Clustering via Low-Rank and Smooth Tensorized Label Learning
Abstract
Tensor-based multi-view clustering with anchor graphs effectively exploits complementary information across views while reducing computational complexity, leading to impressive results. Existing methods typically impose low-rank constraints to capture global consistency, but often overlook the local smoothness prior inherent in real-world data, where neighboring samples on a manifold should share similar attributes. To address this limitation, we propose Multi-View Clustering via Low-Rank and Smooth Tensorized Label Learning (MVC-LSTLL). Our framework unifies low-rank and smoothness objectives through a novel tensor total variation Schatten p-norm with a non-convex surrogate, which simultaneously captures both priors and offers a tighter rank approximation than traditional tensor nuclear norm (TNN). We further propose an adaptive anchor graph learning strategy that dynamically aligns data-anchor relationships with the low-rank smooth representation to enhance clustering accuracy. We develop an optimization approach based on the Augmented Lagrangian Multiplier (ALM) algorithm for the proposed model. Extensive experiments on multiple benchmarks confirm the effectiveness and superiority of our approach.