Continual multi-view clustering with pseudo-label guided consistency anchor learning

Area of Science:

• Machine Learning
• Data Mining
• Artificial Intelligence

Background:

• Continual multi-view clustering (CMVC) addresses challenges in discovering cluster structures from sequential multi-view data with distribution shifts.
• Existing incremental anchor-based CMVC methods suffer from ambiguous anchor semantics, noise sensitivity, and difficulties in aligning anchor graphs, leading to semantic drift and instability.
• These limitations hinder the effective discovery of consistent and stable cluster structures in evolving data.

Purpose of the Study:

• To propose a novel continual multi-view clustering method, CMVC-PCAL, that overcomes the limitations of existing approaches.
• To enhance anchor learning by incorporating pseudo-label guidance for improved semantic clarity and noise robustness.
• To ensure temporal consistency and clustering stability through aligned anchor graphs and refined semantic representations.

Main Methods:

• CMVC-PCAL employs a unified anchor graph representation in initialization, decomposing graphs into shared and view-specific components, and using cluster-wise pseudo labels for alignment.
• A lightweight incremental scheme updates anchors and anchor graphs in the continual stage, utilizing an alignment matrix and consistency regularization.
• Pseudo-label guidance jointly refines anchor semantics and cluster centers, while an alternating optimization algorithm ensures efficient training.

Main Results:

• Experiments on public multi-view benchmark datasets demonstrate the effectiveness of CMVC-PCAL.
• The proposed method shows improved performance in continual multi-view clustering scenarios compared to existing techniques.
• CMVC-PCAL achieves greater robustness to noise and better suppression of semantic drift.

Conclusions:

• CMVC-PCAL effectively addresses the limitations of previous CMVC methods by enhancing anchor learning and graph alignment.
• The pseudo-label guided consistency anchor learning approach leads to more stable and semantically meaningful clustering.
• The method offers an efficient and effective solution for continual multi-view clustering tasks with evolving data distributions.