Contrastive Continual Multiview Clustering With Filtered Structural Fusion.

Area of Science:

• Machine Learning
• Data Mining
• Artificial Intelligence

Background:

• Traditional multiview clustering assumes pre-collected data, failing in sequential real-time scenarios.
• Sequential data collection poses challenges due to privacy and memory constraints, making prior data unavailable.
• Existing methods for sequential data face a stability-plasticity dilemma, leading to catastrophic forgetting of prior knowledge.

Purpose of the Study:

• To propose a novel method, contrastive continual multiview clustering with filtered structural fusion (CCMVC-FSF), for sequential data clustering.
• To address the catastrophic forgetting problem (CFP) in continual learning settings for multiview clustering.
• To enhance the extraction of consistent and complementary information from sequentially arriving data views.

Main Methods:

• Developed a data buffer to store filtered structural information from previous views.
• Utilized contrastive learning to guide the generation of a robust partition matrix using stored information.
• Introduced a 'clustering then sample' strategy to manage the complexity of structural information acquisition and storage.
• Theoretically connected CCMVC-FSF with semisupervised learning and knowledge distillation.

Main Results:

• CCMVC-FSF effectively mitigates catastrophic forgetting in continual multiview clustering.
• The proposed method demonstrates superior performance in clustering sequential data compared to existing approaches.
• The filtered structural fusion and contrastive learning components contribute to robust clustering.

Conclusions:

• CCMVC-FSF offers a promising solution for real-time multiview clustering challenges.
• The method successfully balances plasticity and stability in continual learning scenarios.
• The findings suggest broader applicability in domains requiring sequential data analysis and clustering.