Anchor-based disentanglement framework for incremental multi-view clustering.

Area of Science:

• Machine Learning
• Data Mining
• Artificial Intelligence

Background:

• Incremental Multi-view Clustering (IMvC) methods are gaining traction for handling dynamic data.
• Existing IMvC approaches often suppress unique view characteristics by enforcing a unified representation.
• This limitation hinders effective cross-view knowledge transfer in incremental scenarios.

Purpose of the Study:

• To propose a novel Anchor-based Disentanglement Framework for Incremental Multi-view Clustering (ADIMC).
• To enable conflict-free knowledge transfer in view-incremental scenarios by disentangling representations.
• To preserve both shared semantics and view-specific information for improved clustering.

Main Methods:

• Learning an anchor graph to extract latent semantic information from new views.
• Decomposing the anchor graph into view-consistent and view-specific components.
• Restricting knowledge transfer to view-consistent components while retaining view-specific ones.
• Employing an efficient iterative optimization algorithm for stable model updating.

Main Results:

• ADIMC successfully disentangles shared and private information across views.
• Knowledge transfer is effectively managed, maintaining cross-view coherence and view uniqueness.
• The framework demonstrates stable and efficient model updating with limited computational resources.
• Extensive experiments confirm ADIMC's superiority over state-of-the-art IMvC methods.

Conclusions:

• ADIMC offers an effective solution for incremental multi-view clustering by addressing limitations of unified representations.
• The proposed disentanglement approach facilitates robust knowledge transfer in dynamic, multi-view environments.
• ADIMC provides a superior and efficient method for handling evolving multi-view data.