基于双向注意力的图神经推荐算法研究

A neural network recommender algorithm with bi-directional knowledge graph attention

目的推荐系统主要目标是分析用户的历史行为,以预测用户最感兴趣的项目,随着电子商务和在线服务的日益发展,个性化推荐已成为当今信息传播的基石。基于知识图谱的神经网络推荐通过构建知识图作为输入,可以很好地将节点信息和拓扑结构相结合进行预测,在推荐准确性方面已经证明了良好的结果。然而,现有方法较少考虑图结构中存在的对称关系和信息聚合时梯度消失问题。方法本文提出一种基于知识图谱和神经网络结合的双向注意力机制推荐算法(BGANR),首先将图神经网络与对称注意力机制相结合;然后在不增加额外数据集中维度的情况下,通过双向对称嵌入翻译模型获取用户-项目间的高阶关系,实现对知识图谱中用户-项目信息进行特征的嵌入表示,使注意力机制在决策权重时考虑的关系更全面;其次,基于图神经网络在对节点和邻居信息训练过程中,采用多通道激活函数针对不同的高阶关系进行修正,从而增加反馈的信息量,避免训练过程中的过拟合问题。结果仿真实验结果表明,在Last-FM数据中Recall、NDCG指标与经典模型最好的结果相比,分别提高了2.56%、1.96%。结论BGANR不仅能够实现双向探索高阶连通性,而且在捕捉有效的协同信号的同时能够实现高效信息传递。

Objective Recommendation system,one of the most successful application of e-commerce and online services with the main goal of analyzing a user’s history behavior and then predicting items which are of most interest to users,has become a cornerstone of today’s information dissemination.Comparing with the traditional neural network,the neural network based on knowledge graph(KG)took the building graph as the input in the recommendation system,which could combine the node information and topology for prediction,and had demonstrated good results in terms of recommendation accuracy.However,the existing methods rarely consider the symmetric relationship in the graph structure and the problem of gradient vanishing in information aggregation. Methods A bi-directional graph attention neural network recommendation algorithm (BGANR) was proposed based on the combination of knowledge graph and neural network.Firstly the graph neural network and the symmetric attention mechanism were combined.Then,without adding additional dataset dimensions,the higher-order relationships between users-items were obtained through a bidirectional symmetric embedded translation model,which aimd at embedding representations of the features of user-item information in the Knowledge Graph,so that the relationships were considered by the attention mechanism in the decision-making weights more comprehensively.The graph-based neural network was used to correct different higher-order relationships by using multi-channel activation functions during the training process of node and neighbor information,so as to increase the amount of feedback information and avoid the over-fitting in the training process. Results The Recall and NDCG metrics in Last-FM data were improved by 2.56% and 1.96% respectively,compared with the best results of state-of-the-art model. Conclusion The extensive empirical results demonstrated that BGANR could not only explore the higher-order connectivity in bi-directions,but also realize the efficient transmission of information while capturing effective collaborative signals.