序列推荐可以根据用户与项目之间的历史交互记录来建立用户行为序列,从而动态地建模用户偏好.在序列推荐中,大多数基于深度学习的模型,将用户行为编码成单个向量,但是实际的用户兴趣偏好通常是多个方面的,现有模型不足以捕捉用户兴趣的变化性质.为解决这个问题,文章提出了一种基于时间卷积网络和动态路由的多兴趣推荐模型(Multi-Interest Sequential Recommendation Model Based on Temporal Convolutional Network and Dynamic Routing,MIRTD).文章先使用基于胶囊路由机制的多兴趣提取器层,聚类历史行为和提取不同的兴趣,然后利用时间卷积网络进一步捕捉不同兴趣序列,从而提取到更全面的特征.实验结果表明,相较于较为流行的序列推荐模型,MIRTD在点击率等指标上展现出更加优越的性能.
Multi-Interest Sequential Recommendation Model Based on Temporal Convolutional Network and Dynamic Routing
Sequential recommendation can dynamically model user preferences by building user behavior sequences based on historical interaction records between users and projects.In sequential recommendation,most existing deep learning-based models encode user behavior into a single vector,but users'interest preferences are usually multi-faceted,and the existing models are insufficient to capture the changing nature of users'interests.To solve these problems,this paper proposes a Multi-Interest Recommendation Model based on Temporal Convolutional Network and Dynamic Routing(MIRTD).First,we use a multi-interest extractor layer based on a capsule routing mechanism to cluster historical behaviors and extract different interests.Then,this paper uses the time convolutional network to further capture different interest sequences,so as to extract more comprehensive features.The experimental results show that MIRTD has better performance in click through rate and other indicators than the more popular sequential recommendation models.
万方数据
维普
