AdaptCMVC++: Robust and Flexible Adaptation to Incremental Views in Continual Multi-view Clustering.

Abstract

Current multi-view clustering methods are developed under the assumption that all views are simultaneously accessible to the model. However, this assumption can break down in real-world scenarios where views are incrementally acquired over time, necessitating the development of continual multi-view clustering (CMVC) approaches. Existing CMVC methods typically adopt late-fusion strategies, training a separate model for each incoming view to extract view-specific information-such as partition matrices, similarity matrices, or latent representations-which are then used to update a shared consensus representation via a moving average mechanism. However, these methods are sensitive to view-specific noise and struggle to handle large discrepancies across views. To address these limitations, we revisit CMVC from a domain adaptation perspective and propose AdaptCMVC++, which continuously integrates information from newly available views while mitigating catastrophic forgetting. Specifically, a self-training framework is introduced to extend the model to new views, specifically designed to be robust to view-specific noise. To combat catastrophic forgetting, a structure-alignment mechanism is proposed to enable the model to explore the global group structure across multiple views. Furthermore, a dimensionality adaptation module is incorporated to accommodate multi-view data with diverse image dimensionalities. Extensive experiments on several multi-view benchmarks and a newly constructed dataset demonstrate the effectiveness and generalization capability of our proposed method for the CMVC task. The implementation is publicly available at: https://github.com/Wjing-bjtu/AdaptCMVC AdaptCMVC++.