Multi-branch Space Sharing Feature Aggregation for contrastive multi-view clustering

Yuanyang Zhang, Weiqing Yan, Chang Tang, Wujie Zhou, Jian Jin

Abstract: Highlights•We propose a novel method called Multi-branch Space Sharing Feature Aggregation for Contrastive Multi-view Clustering, which achieves sample-wise enhancement by learning sharing sample representation space for different batches data, and achieves end-to-end clustering.•We design a multi-branch space sharing aggregation module to filter out view-specific private information and avoid local sample-wise fusion. This module first maps samples from different batches into a low-dimensional shared representation space, calculates the weights of the samples across different dimensions in this space using nonlinear mapping functions, and finally reconstructs the sample representations based on these weights.•We design a dual cross-view contrastive learning module that integrates the learned consensus representations into instance-wise feature contrastive learning and category-wise semantic label contrastive learning. This aligns features from different views and captures the consistency of semantic information, thereby further enhancing the separability and compactness of the end-to-end clustering.

External IDs:dblp:journals/pr/ZhangYTZJ25