Dynamic Multi-View Learning With Multi-System State Space Model

As real-world data become more heterogeneous, multi-view learning has emerged as an increasingly vital research area. Even though most existing methods adopt graph convolutional networks to model multi-view structures and features with remarkable success, this paradigm is inherently static in capturing view data dependencies and lacks dynamic adaptability to downstream changes. The emergence of state space models explores the dynamics of state transitions in the hidden space, capturing temporal dependencies and evolving patterns that mitigate the above phenomenon. However, existing state space models are always limited in their ability to explore relationships across multiple modalities. Therefore, we propose a method named Dynamic Multi-view State Space Model (DMvSSM) that explores the evolution of view states and the interactions among different views, while maintaining low time complexity. Specifically, we dynamically explain the multi-view learning by constructing a multi-system state space model, which incorporates topology and semantic space. In the topology space, we establish attention-based topology dependencies to effectively uncover the structural patterns of each view. In the semantic space, we develop a Hebbian-based cross-view diffusion equation that extracts globally expressive features and facilitates selective cross-view interactions.The empirical research on several benchmark datasets substantiates the benefits of the proposed method and performs effectively with large-scale datasets.