基于双通道粒计算的深度多视图聚类方法OA

In order to deal with the problems of quality differences among different views,ambiguous boundary samples and differences in semantic structures among different views,we propose deep multi-view clustering with dual-channel granular computing.A dual-channel feature fusion module is designed to strengthen key representations,where the global average pooling channel captures holistic semantics,and the global max pooling channel focuses on highly discriminative cues.Furthermore,a dual-channel contrast learning strategy is introduced for contrast learning at the sample and local fuzzy granular-ball structure level respectively.Fuzzy granular-ball level contrast learning is divided into intra-granular-ball and cross-view fuzzy granular-ball contrast learning.The former optimizes the clustering boundary by making positive samples inside the granular-ball closer.The latter ensures consistent granular-ball structures are learned across different views.Additionally,this paper introduces a view-adaptive attention weight assignment mechanism that enhances the leading role of high-quality views in clustering.We verify the effectiveness of our method on eight publicly available multi-view datasets.The results show that our method improves clustering accuracy compared to the existing multi-view clustering methods,such as MFLVC,SCMVC,etc.