多兴趣点融合多模态知识图谱的跨会话推荐  

Cross-session recommendation based on fusion of multiple interest points and multimodal knowledge graphs

作　　者：陈刚 孙伟 张丽英 陈平华[3] CHEN Gang;SUN Wei;ZHANG Li-ying;CHEN Ping-hua(School of Artificial Intelligence,Guangdong Open University,Guangzhou 510091,China;School of Computer Science and Engineering,Guangzhou Institute of Science and Technology,Guangzhou 510540,China;School of Computer Science and Technology,Guangdong University of Technology,Guangzhou 510006,China)

机构地区：[1]广东开放大学人工智能学院,广东广州510091 [2]广州理工学院计算机科学与工程学院,广东广州510540 [3]广东工业大学计算机学院,广东广州510006

出　　处：《计算机工程与设计》2024年第12期3749-3757,共9页Computer Engineering and Design

基　　金：广东省重点领域研发计划基金项目(2020B0101100001);广东省软科学研究计划基金项目(2024A1010010001)。

摘　　要：针对现有的会话推荐专注单一兴趣的上下文,忽略了单用户多兴趣,以及利用知识图谱时忽略多模态数据类型的不足,提出一种多兴趣点融合多模态知识图谱的跨会话推荐。受少样本学习在有限实例学习模型的启发,设计一个跨会话协作网络,将下一个项目推荐建模为少样本学习问题;从用户的行为序列中捕获用户的各种兴趣,根据用户的历史和当前行为序列构建兴趣图;引入多模态知识图注意力网络,通过使用多模态图注意力机制进行信息传播,得到聚集嵌入表示并进行推荐;设计一个相似会话检索网络,从历史会话中找出与当前会话相似的网络来补充和优化偏好表示。实验结果表明,所提算法在Recall@20和MRR@20指标上均优于基线。Aiming to address the limitations of current session-based recommendation methods,which primarily only focus on the context of a single interest and neglect the multiple interests of a single user,and overlook the utilization of multimodal data types when incorporating knowledge graphs,a cross-session recommendation algorithm based on multi-interest point fusion and multi-modal knowledge graph was proposed.Inspired by the successful application of few-shot learning in limited instance lear-ning models,a cross-session collaboration network was designed and the next item recommendation was modeled as a few-shot learning problem.Various interests of users were captured from their behavior sequences and an interest graph based on their historical and current behavior sequences was constructed.The modal knowledge graph attention network was introduced,by using the multi-modal graph attention mechanism,the information was disseminated,aggregated embedding representation was obtained and recommendations were made.A similar session retrieval network was designed to find networks similar to the current session from historical sessions to supplement and optimize the preference representation.Experimental results indicate that the proposed algorithm exhibits excellent performance in terms of Recall@20 and MRR@20 compared to the baseline.

关 键 词：会话推荐 兴趣点 知识图谱 少样本学习 多模态 图注意力 用户偏好 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]