基于图神经网络的异构信任推荐算法

Heterogeneous Trust Recommendation Algorithm Based on Graph Neural Networks

传统基于图神经网络的社交推荐算法通过加强用户和项目特征的学习提升预测精度,但随着用户数据日益稀疏和社交关系趋于复杂,推荐质量提升缓慢。为挖掘用户和项目的潜在关联关系,提出一种结合图神经网络的异构信任推荐算法(GraphTrust)。在显式信任关系的基础上获取用户的潜在好友,根据动态影响力传播模型将图神经网络中的节点和边进行分类,通过不同类型的边在不同节点间进行影响力传播扩散,捕捉隐藏在高阶网络结构中的影响力扩散特征,并使用户和项目的潜在特征随着影响力传播过程达到平衡状态,最终将用户交互的项目特征作为辅助特征与用户特征聚合进行评分预测。在Yelp和Flickr数据集上的实验结果表明,当潜在特征维数为64时,GraphTrust算法相比于DiffNet++算法的命中率和归一化折损累计增益分别提升了13.2%、22.2%和20.4%、25.5%,在一定程度上提高了推荐过程的可解释性和预测精度,并且缓解了数据稀疏问题。

Traditional social recommendation algorithms based on graph neural networks improve prediction accuracy by enhancing the learning of user and item features;however,the recommendation quality slowly decreases as the data about users become increasingly sparse,and social relationships tend to be complex.To explore the potential relationship between users and items,this study proposes a heterogeneous trust recommendation algorithm combining graph neural networks(GraphTrust). The algorithm obtains potential friends of users on the basis of explicit trust relationships.According to the dynamic influence propagation model,the nodes and edges in a graph neural network are classified,and influence propagation and diffusion are carried out between different nodes through different types of edges,to capture the influence diffusion features hidden in the high-order network structure. Additionally,this is to enable the potential characteristics of users and items to reach a balance with the influence dissemination process.Finally,the item features interacted with by users are aggregated as auxiliary and user features for scoring prediction. The experimental results show that when the potential feature dimension is 64,the Hit Rate(HR) of GraphTrust on Yelp and Flickr datasets is increased by 13.2%,22.2% compared with a neural influence and interest diffusion network for social recommendation(DiffNet++),and the Normalized Discounted Cumulative Gain(NDCG)of GraphTrust is increased by20.4%,25.5%,which improves the interpretability and prediction accuracy of the recommendation process to a certain extent,and alleviates the problem of data sparsity.