Dynamically adaptive negative pairs for contrastive multiview clustering

Recently, the issue of handling view-unaligned multi-view clustering has attracted widespread attention. Most current multi-view clustering methods for view-unaligned data attempt to learn invariant representations via contrastive learning. However, prior methods either cannot consider false negative pairs or simply treat a fixed proportion of negative samples as false negative pairs. To address this issue, we propose a novel contrastive multiview clustering via dividing dynamically negative pairs. Specifically, we design a k-means-based strategy to determine a robust contrastive loss, enabling dynamic partitioning of negative pairs. Subsequently, we recover the view-unaligned multi-view data to facilitate clustering. Finally, we perform experiments to show the effectiveness and robustness of our method.