基于双重注意力机制的MIND微博推荐算法  

MIND Microblogging Recommendation Algorithm Based on Dual Attention Mechanism

作　　者：彭丹 魏嘉银 卢友军 姚林 王倩 PENG Dan;WEI Jiayin;LU Youjun;YAO Lin;WANG Qian(School of Data Science and Information Engineering,Guizhou Minzu University,Guiyang 550025,China;Personnel Office,Guizhou Minzu University,Guiyang 550025,China)

机构地区：[1]贵州民族大学数据科学与信息工程学院,贵阳550025 [2]贵州民族大学人事处,贵阳550025

出　　处：《湖北民族大学学报（自然科学版）》2024年第4期507-513,共7页Journal of Hubei Minzu University:Natural Science Edition

基　　金：贵州省科技计划项目(黔科合基础[2018]1082号,黔科合基础[2019]1159号);贵州省教育厅自然科学研究项目(黔教技[2022]015号,黔教技[2022]047号,黔教技[2023]012号,黔教技[2023]061号);贵州民族大学基金科研项目(GZMUZK[2023]YB13)。

摘　　要：针对目前在微博推荐领域主要使用单一向量表示用户兴趣且缺乏对兴趣之间复杂关系的捕捉能力,导致用户兴趣表示不全面,推荐准确性较低的问题,提出了基于双重注意力机制的多兴趣动态路由微博推荐算法(multi-interest network with dynamic routing microblogging recommendation algorithm based on dual attention mechanism,MINDDouAtt),用于提高用户兴趣的表征能力。首先,通过动态路由从用户行为数据中提取多个兴趣胶囊,并将这些兴趣胶囊输入到自注意力机制中以对不同兴趣胶囊之间的关联信息进行交叉学习,提高兴趣的表征能力。然后,通过引入标签感知注意力机制来调节不同兴趣胶囊之间的重要性,以更好地满足用户的个性化推荐需求。实验表明,MINDDouAtt算法在亚马逊图书、天猫和微博数据集上的S HR@10值相较于最好的对比模型分别提升了33.66%、10.49%、9.60%。该算法能够在电子商务等领域为用户提供更准确和个性化的推荐结果。Aiming at the current problem that in the microblogging recommendation field a single vector was mainly used to represent user interests and capture ability of the complex relationship between interests was lacking,which led to incomplete representation of user interests and low recommendation accuracy,a multi-interest network with dynamic routing microblogging recommendation algorithm based on dual attention mechanism(MINDDouAtt)was proposed to improve the characterization ability of user interests.It was proposed for improving the characterization of user interests.First,multiple interest capsules were extracted from user behavioral data through dynamic routing,and these interest capsules were fed into the self-attention mechanism to cross-learn the association information between different interest capsules to improve the characterization ability of interests.Then,the importance between different interest capsules was adjusted by introducing a label-aware attention mechanism to better meet users′personalized recommendation needs.The experiment showed that the proposed model performed well on Amazon Books,Tmall,and Weibo datasets.Compared with the best comparison model,the values of S HR@10 improved by 33.66%,10.49%,and 9.60%respectively.The MINDDouAtt algorithm can provide users with more accurate and personalized recommendation results in fields such as e-commerce.

关 键 词：微博推荐 多兴趣召回 动态路由 兴趣胶囊 自注意力机制 标签感知注意力机制 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]