VDMN: View-Driven Modular Network for Consistency and Complementarity Mining in Multi-View Clustering

Multi-view clustering, which leverages the complementarity and consistency among diverse feature sets to enhance clustering performance, is attracting growing attention from researchers for its broad applicability in fields such as pattern recognition and data mining. Traditional approaches typically explore this characteristic by mapping different views into a unified embedding space using view-specific linear or nonlinear neural networks. Subsequently, this unified embedding serves as input for standard single-view clustering algorithms to derive the final clustering outcomes. Nevertheless, relying solely on a unified embedding may be inadequate for capturing the unique or conflicting attributes of multi-view data, given their varied representation capabilities. To alleviate the above problem, a view-driven modular network is proposed to adaptively model the relationship between different views, where base network is used for all views and view-specific routing network is constructed for different views. Furthermore, an objective function grounded in Kullback-Leibler divergence is introduced as a clusteringoriented constraint, facilitating the learning of embeddings that are more conducive to effective clustering. Extensive empirical evaluations across six benchmark datasets have demonstrated superior performance when compared to several state-of-the-art methods.