融合协同效应的自适应图卷积网络推荐算法

Adaptive graph convolutional network recommendation algorithm incorporating collaborative effects

基于图神经网络的协同过滤算法在许多推荐场景中取得了良好的表现,但是现有的图神经网络推荐模型在推荐过程中忽略了节点间协同效应与节点-层级的重要性融合表达以及堆叠多层网络容易存在过平滑问题。为解决上述问题,文中提出一种新的融合协同效应的自适应图卷积网络推荐算法。首先,构建用户-项目二部图并计算节点间交互率;其次,根据节点局部结构自适应决定节点-层级的融合表达权重;然后,利用第l层与第l+1层嵌入向量的相似度来缓解图卷积网络的过平滑问题;最后,使用协同效应融合自适应图卷积网络学习潜在的特征并通过内积作为结果进行推荐。在五个真实数据集上的实验结果表明,所提模型相比基线模型,在召回率和归一化折损累计增益两个评价指标上有明显提高。

The collaborative filtering(CF)algorithms based on graph neural networks(GNNs)have shown promising performance in many recommendation scenarios.However,the existing GNN recommendation models overlook the importance fusion expression of collaborative effects among nodes and the node⁃hierarchy,as well as the potential over⁃smoothing inherent in stacking multiple layers of networks.In view of the above,a novel adaptive graph convolutional network(GCN)recommendation algorithm that integrates collaborative effects is proposed.A bipartite graph of users and items is constructed and the interaction rate among nodes is calculated.The fusion expression weight of node⁃hierarchy is determined adaptively based on the local structure of nodes.The over⁃smoothing of GCNs is mitigated by measuring the similarity between embedding vectors of the lth and(l+1)th layers.The collaborative effects fused with adaptive GCNs are used to learn the latent features and make recommendation by taking inner products as the results.The results of experiments on five real datasets demonstrate that the proposed model outperforms the baseline models in terms of the evaluation indexes of recall rate and normalized discounted cumulative gain(NDCG).