基于多子图的图卷积网络推荐模型

Recommendation Model of Graph Convolutional Network Based on Multi-Subgraph

摘要: 随着深度学习的崛起，将图卷积网络应用到推荐模型中，虽然带来了显著的性能提升，但同样面临着过平滑问题——随着网络层数的增加，所有节点的嵌入表示变得相似，失去了原有的差异化信息，导致模型的表征能力下降，影响推荐性能。本文提出一种基于多子图的图卷积网络推荐模型(MSGCN)，该模型在子图中进行高阶嵌入传播。子图由具有相似偏好的用户及其交互的物品组成。为了生成子图，采用图注意力网络对用户节点进行分类并划分到相应的子图中。改进后的模型避免了在高阶嵌入传播层中将差异显著的节点强行关联，缓解了图卷积网络模型在训练过程中常遇到的过平滑现象。通过在三个数据集上实验来验证MSGCN的有效性。实验结果表明，MSGCN模型相较于最优模型在Recall@20分别提升4.51%，6.69%和10.61%，在NDCG@20分别提升5.89%，10.93%和11.71%。

Abstract: With the rise of deep learning, applying graph convolutional networks to recommendation models has brought significant performance improvements, but it also faces the problem of over-smoothing —as the number of network layers increases, the embedding representations of all nodes become similar, losing the original differentiated information, resulting in a decline in the model’s representation ability and affecting the recommendation performance. This paper proposes a recommendation model of graph convolutional network based on multi-subgraph (MSGCN). The model performs high-order embedding propagation within the subgraph. The subgraphs are composed of users with similar preferences and the items they interact with. To generate subgraphs, a graph attention network is used to classify user nodes and divide them into corresponding subgraphs. The improved model avoids forcibly associating nodes with significant differences in the high-order embedding propagation layer, alleviating the over-smoothing phenomenon that graph convolutional network models often encounter during training. The effectiveness of MSGCN is verified by experiments on three datasets. The experimental results show that the MSGCN model outperforms the best model in Recall@20 by 4.51%, 6.69%, and 10.61%, and in NDCG@20 by 5.89%, 10.93%, and 11.71%.