Topology-Driven Multi-View Clustering Via Tensorial Refined Sigmoid Rank Minimization

Zhibin Gu,Zhendong Li,Songhe Feng
DOI: https://doi.org/10.1145/3637528.3672070
2024-01-01
Abstract:Benefiting from the effective exploitation of the high-order correlations across multiple views, tensor-based multi-view clustering (TMVC) has garnered considerable attention in recent years. Nevertheless, prior TMVC techniques commonly involve assembling multiple view-specific spatial similarity graphs into a three-dimensional tensor, overlooking the intrinsic topological structure essential for precise clustering of data within a manifold. Additionally, mainstream techniques are constrained by equally shrinking all singular values to recover a low-rank tensor, limiting their capacity to distinguish significant variations among different singular values. In this investigation, we present an innovative TMVC framework termed toPology-driven multi-view clustering viA refined teNsorial sigmoiD rAnk minimization (PANDA ). Specifically, PANDA extracts view-specific topological structures from Euclidean graphs and intricately integrates them into a low-rank three-dimensional tensor, facilitating the concurrent utilization of intra-view topological connectivity and inter-view high-order correlations. Moreover, we develop a refined sigmoid function as the tighter surrogate to tensor rank, enabling the exploration of significant information of heterogeneous singular values. Meanwhile, the topological structures are merged into a unified structure with varying weights, associated with a connectivity constraint, empowering the significant divergence among views and the explicit cluster structure of the target graph are simultaneously leveraged. Extensive experiments demonstrate the superiority of PANDA, outperforming SOTA methods.
What problem does this paper attempt to address?