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

The core of partially view-aligned clustering is to address the issue of instance misalignment that occurs between different views during data collection. However, existing methods rely excessively on predefined alignment samples and are prone to failure when there is complete view misalignment. Inspired by game theory, we propose a Banzhaf game theory-driven cross-view Interaction aligNment method, dubbed BIN, which fundamentally mitigates the reliance on pre-aligned samples and enables robust clustering even under severe view misalignment. This approach treats multi-view samples as game players, using the Banzhaf index to quantify the marginal contributions of alliance members, accurately modeling complex sample relationships. Simultaneously, a dual loss constraint mechanism is designed: the Banzhaf gain loss dynamically captures the marginal contributions of cross-view sample pairs, enhancing sample correlation, while the contrastive loss suppresses interference from weakly correlated samples, constructing a mechanism that can autonomously learn sample correspondences without pre-alignment. Furthermore, a cross-view fusion reconstruction representation strategy is introduced, which adaptively fuses information and eliminates irrelevant noise interference, ensuring that the Banzhaf index operates accurately in a representation space with high semantic relevance and low bias. We conducted extensive and in-depth experiments on multi-view data, and the experimental results demonstrate the superior performance of our proposed method, strongly validating its effectiveness and practicality.