主题：Recommendation Systems Leveraging Multi-view Graph Contrastive Learning and Online Distillation

基于多视角图对比学习与在线蒸馏的推荐系统算法研究

时间: 2025年12月10日 上午10:30

地点：管理科研楼 第一教室

主讲人：寇纲 教授

Bio: 寇纲现任全国政协委员、湘江实验室副主任，西南财经大学大数据研究院院长、中国系统工程学会副理事长、长江学者特聘教授、国家杰出青年科学基金获得者、国务院享受政府特殊津贴专家。主持社科重大等多项科研课题；在Science，Nature子刊，UTD24期刊（ISR, JOC）和ICML、AAAI、KDD等顶会发表200余篇论文，H指数77，论文被他人引用2万余次。以第一完成人身份获教育部高等学校科学研究优秀成果奖自然科学一等奖、人文社会科学一等奖等多项省部级科研奖励，所撰写的10余份政策建议曾获得习近平总书记等中央领导人批示。

Abstract: Recommendation systems have become deeply embedded in people's daily lives, and users' heavy reliance on them poses non-negligible potential risks to mental health. Furthermore, in healthcare applications, recommendation technology has been successfully deployed across multiple domains including disease prediction, prevention, and medical diagnosis, demonstrating significant value. Typically, recommendation systems leverage rich historical interaction data between users and items to achieve accurate recommendations. However, in practical applications, new users or items often face the cold-start problem due to lack of interaction data. Simultaneously, insufficient interaction information leads to data sparsity issues. These two challenges severely constrain the performance of recommendation systems. To address these limitations, this study first proposes a multi-view graph contrastive learning approach that jointly models attributes and structure. Through an adaptive contrastive learning module, our method dynamically regulates the mutual information levels between the final contrastive views, thereby fully exploiting the rich information contained in both attribute and structural views. For the data sparsity problem, we propose a multi-view fusion recommendation framework. This framework utilizes multi-view graph contrastive learning to integrate user social relationships and item semantic associations, effectively mitigating the negative impact of data scarcity. For cold-start scenarios, we design a bidirectional online distillation mechanism. This enables two-way knowledge transfer between the "content-enhanced collaborative embedding network" and the "content-based embedding network," achieving adaptive fusion of content information and collaborative signals. This approach effectively resolves the cold-start problem while enhancing recommendation performance.