融合元路径与改进协同注意力的推荐研究  

Research on Recommendation Fusing Meta-Path and Improved Collaborative Attention

作　　者：秦倩倩 张志军[1] 袁卫华[1] 桑胜 胡海星 王宜贵 QIN Qian-qian;ZHANG Zhi-jun;YUAN Wei-hua;SANG Sheng;HU Hai-xing;WANG Yi-gui(School of Computer Science and Technology,Shandong Jianzhu University,Jinan 250101,China)

机构地区：[1]山东建筑大学计算机科学与技术学院,山东济南250101

出　　处：《计算机技术与发展》2022年第12期150-158,共9页Computer Technology and Development

基　　金：国家自然科学基金项目(61902221);山东省自然科学基金项目(ZR2021MF099);山东省教学改革研究项目(M2021130)。

摘　　要：现有的基于异质信息网络的推荐方法通过最大池化操作提取元路径语义信息时,未考虑元路径语义的整体性特征及特征冗余问题的影响。针对这些问题,提出了一种融合元路径与改进协同注意力的推荐模型MICA(Research on Recommendation Fusing Meta-Path and Improved Collaborative Attention)。该模型通过基于k-最大池化的协同注意力机制更深层次地学习用户和项目的协同注意力嵌入表达,缓解了最大池化操作在协同注意力机制中带来的特征丢失问题;基于提取的路径序列显著性特征和完整性特征通过注意力机制学习元路径的重要特征,得到元路径上下文嵌入表达,保留有效的元路径语义信息;利用注意力机制学习融合后的用户、项目协同注意力嵌入表达和元路径上下文嵌入表达,以减少冗余信息,实现top-N推荐。在两个真实数据集上的实验结果表明,MICA在三个评价指标上均优于其他模型,能够更好地提高推荐性能,有效地提取特征信息,缓解元路径特征提取不充分等问题。When the existing recommendation methods based on heterogeneous information network extract meta-path semantic information by maximum pooling operation,the influence of the overall characteristics of meta-path semantics and feature redundancy cannot be taken into account.To solve these problems,we propose a recommendation model named MICA,which fuses meta-path and improved collaborative attention mechanism.Firstly,MICA uses a collaborative attention mechanism based on k-max pooling to learn the collaborative attention embedding representations of users and items in a deeper level,respectively,which alleviates the problem of feature losses caused by max pooling in the collaborative attention mechanism.Secondly,based on the significant feature and integral feature,the meta-path context embedding representation is obtained through the attention mechanism,a mechanism to learn important feature of meta-path,to retain effective meta-path information.Lastly,to reduce redundant information,the user and item collaborative attention embedding representations and meta-path context embedding representations are fused to input into attention mechanism for top-N recommendations.Experiments on two real datasets show that MICA is superior to other state-of-the-art models in three evaluation indices,which better improves the recommendation performance.The model can effectively extract feature information and alleviate the problem of insufficient feature extraction of meta-path.

关 键 词：异质信息网络 推荐系统 协同注意力机制 元路径 最大池化 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]