基于异质信息网络与图注意力的深度学习推荐算法研究

Research and Implementation of Deep Learning Recommendation Algorithm based on Heterogeneous Information Network and Graph Attention

随着个性化推荐系统旨在精准捕捉用户偏好以做出合理的推荐。以往的个性化推荐算法大多基于协同过滤等方式建模,近年来采用的机器学习方法虽在传统方法基础上取得了较大的进步,但当前现有方式普遍只考虑了用户与内容间单行为及内容与内容间单关系,无法有效地从用户与内容多种交互行为关系中提炼复杂协同信号。为解决该问题,提出了一个基于图注意力机制和异质信息网络的个性化推荐算法(HA-Rec)。首先,通过构建用户与内容交互多行为与内容之间多关系的异质图,实现信息的最大化保留;此外,通过基于异质图注意力机制对内容侧图谱进行处理以从中挖掘有效特征,并将交互时间信息、用户信息及内容信息进行全面融合,实现用户-内容侧复杂关系的捕捉与节点嵌入信息表示。最后,通过在公开数据集上进行对比实验验证及消融实验表明,所提出的模型在个性化推荐任务中取得了更好的效果。

The personalized recommendation system aims to accurately capture user preferences to make reasonable recommendations.Most of the previous personalized recommendation algorithms are based on collaborative filtering.Although the machine learning methods adopted in recent years have made great progress on the basis of the traditional methods,the current existing methods generally only consider the single behavior between users and content and the single relationship between content and content,and cannot effectively extract complex collaborative signals from the multiple interactive behavior relationships between users and content.To solve this problem,this paper proposes a personalized recommendation algorithm based on graph attention mechanism and heterogeneous information network(HA-Rec).First of all,we can maximize the retention of information by constructing a heterogeneous graph of multiple behaviors between users and content and multiple relationships between content;In addition,the content side atlas is processed based on the heterogeneous graph attention mechanism to mine effective features from it,and the interaction time information,user information and content information are comprehensively integrated to achieve the capture of user-content side complex relationship and node embedded information representation.Finally,the comparison and ablation experiments on the public dataset show that the model proposed in this paper achieves better results in personalized recommendation tasks.