融合知识图谱与注意力机制的个性化序列推荐  

Personalized Sequence Recommendation Algorithm Based on Knowledge Graph and Attention Mechanism

作　　者：任永功[1] 吕福泽 张志鹏 REN Yong-gong;LV Fu-ze;ZHANG Zhi-peng(School of Computer and Information Technology,Liaoning Normal University,Dalian 116000,China)

机构地区：[1]辽宁师范大学计算机与信息技术学院,辽宁大连116000

出　　处：《小型微型计算机系统》2022年第7期1362-1369,共8页Journal of Chinese Computer Systems

基　　金：国家自然科学基金项目(61976109)资助;辽宁省“兴辽英才计划”项目(XLYC2006005)资助;中国博士后科学基金项目(2021M690527)资助;辽宁省自然科学基金博士启动项目(2020-BS-184)资助.

摘　　要：序列推荐系统可以根据用户和物品交互的时间序列信息,精确预测用户下一次交互物品.现有的序列推荐算法存在用户兴趣过渡拟合的问题,导致推荐内容同质化严重,从而无法实现个性化推荐.基于此,本文提出一种融合知识图谱与注意力机制的个性化序列推荐算法(SR-KGA):首先,引入知识图谱,通过图卷积网络对物品进行嵌入表示;其次,通过自注意力机制和多头注意力机制构建序列到序列(seq2seq)模型,最后,在损失函数中加入多样性正则项;实现用交互序列来预测未来可能交互的物品序列,从而进行推荐.通过在真实的数据集上实验,SR-KGA在保证推荐准确度的同时,提升了推荐列表的多样性,实现了用户个性化推荐.Sequence recommender system can utilize the time sequence interaction information between users and items to predict a target user′s next interactive item accurately.However,existing sequence recommendation algorithms suffer from user interest overfitting problem,resulting in serious homogeneity of recommended content and cannot provide personalized recommendations.To this end,a personalized sequence recommendation algorithm that integrates knowledge graph and attention mechanism is proposed in this paper.Firstly,knowledge graph is employed to embed the sequence by graph convolutional network.Then,sequence to sequence model is constructed by self-attention mechanism and multi-attention mechanism.Finally,add a diversity regularized to the loss function;historical interaction sequence is utilized to predict and recommend the sequence of items that the targe user may prefer.Experiments executed on real-world datasets suggest that the proposed approach can realize personalized recommendations can improve the diversity of the recommendation list while ensuring the recommendation accuracy,so as to realize the personalized recommendations.

关 键 词：序列化推荐系统 知识图谱 注意力机制 个性化推荐 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]