Game Theory Inspired Cross-View Interaction Alignment for Partially View-Aligned Clustering.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Instance misalignment across data views is a key challenge in multi-view clustering.
• Current methods often fail due to over-reliance on predefined alignment samples, especially with complete view misalignment.

Purpose of the Study:

• To develop a robust cross-view alignment method for multi-view clustering that mitigates reliance on pre-aligned samples.
• To enhance clustering performance even under severe view misalignment conditions.

Main Methods:

• Proposed a Banzhaf game theory-driven cross-view interaction alignment (BIN) method.
• Utilized the Banzhaf index to model complex sample relationships by quantifying marginal contributions.
• Implemented a dual loss constraint mechanism (Banzhaf gain loss and contrastive loss) for autonomous sample correspondence learning.
• Introduced a cross-view fusion reconstruction strategy for adaptive information fusion and noise reduction.

Main Results:

• The BIN method demonstrated superior performance in multi-view clustering tasks.
• Effectively mitigated the need for pre-aligned samples, enabling robust clustering under severe misalignment.
• Experimental results validated the method's effectiveness and practicality on diverse multi-view datasets.

Conclusions:

• The proposed Banzhaf game theory-driven approach offers a significant advancement in multi-view clustering.
• BIN provides a robust and practical solution for handling instance misalignment without extensive pre-alignment.
• The method shows strong potential for applications requiring accurate analysis of multi-view data.