Recently, the acceptance results of the 34th ACM International Conference on Multimedia (ACM MM 2026) have been announced. A research paper titled Global-and-Local Collaborative Mixture-of-Experts with Dual Contrastive Learning for Deep Multi-view Clustering, authored by the research team led by Professor Wang Youqing from our school, has been successfully accepted.
Professor Wang Youqing (graduate supervisor) serves as the first author; Xiao Bin, a 2023 master’s student from the School of Information Science and Technology, is the second author; Associate Professor Guo Jipeng (co-supervisor) acts as the corresponding author. Beijing University of Chemical Technology is the primary affiliation and the sole corresponding institution of this work.
To address the drawbacks of existing deep multi-view clustering methods—namely, their neglect of cross-view heterogeneity and difficulties in semantic alignment across different views—this study proposes a novel Global-and-Local Collaborative Mixture-of-Experts (GLCMoE) framework. The framework consists of three core modules: Local Mixture-of-Experts (LMoE), Global Mixture-of-Experts (GMoE), and Dual Contrastive Learning (DCL).
The LMoE module leverages a parameter-shared expert bank to filter cross-view heterogeneity. With a gating routing mechanism that adaptively activates and reorganizes collaborative experts, it preserves complementary fine-grained view-specific information.The GMoE module adopts Transformer attention-based routing. It focuses on extracting cross-view consistency from a unified multi-view feature space and capturing high-order cross-view interactions.The DCL module achieves deep semantic alignment across views at multiple levels: coarse-grained alignment between local view-specific representations, and fine-grained alignment between fused local representations and global representations.Additionally, a gating routing equilibrium regularization term is designed to mitigate expert collapse and boost the utilization rate of all experts. Following the strategy of "split first, aggregate later, and conduct hierarchical contrastive learning", the GLCMoE method realizes discriminative cross-view representation fusion and semantic alignment.
The ACM International Conference on Multimedia (ACM MM) is a flagship international academic conference hosted by the Association for Computing Machinery (ACM). Founded in 1993, it has long stood as the premier global conference for multimedia research, innovation and applications, as well as one of the most influential top-tier international conferences in multimedia processing, analysis and computing. On the China Computer Federation (CCF) conference ranking list, ACM MM is categorized as the sole Category A international conference in the field of computer graphics and multimedia. The acceptance of this paper marks another breakthrough achievement for our school at the world’s top-tier academic conference focused on computer multimedia.
This research was supported by the National Natural Science Foundation of China and the Beijing Natural Science Foundation, among other funding programs.