ISSN 1000-1239 CN 11-1777/TP

计算机研究与发展 ›› 2021, Vol. 58 ›› Issue (9): 1987-1996.doi: 10.7544/issn1000-1239.2021.20200617

• 人工智能 • 上一篇    下一篇



  1. (西南财经大学经济信息工程学院 成都 610074) (
  • 出版日期: 2021-09-01
  • 基金资助: 

A Collaborative Recommendation Model Based on Enhanced Graph Convolutional Neural Network

Wang Lei, Xiong Yuning, Li Yunpeng, Liu Yuanyuan   

  1. (School of Economics Information Engineering, Southwestern University of Finance and Economics, Chengdu 610074)
  • Online: 2021-09-01
  • Supported by: 
    This work was supported by the Research Planning Fund of Humanities and Social Sciences of the Ministry of Education (16XJAZH002) and the Fundamental Research Funds for the Central Universities (JBK2102049).

摘要: 图卷积神经网络是一种针对图结构数据的深度学习模型,由于具有强大的特征提取和表示学习能力,它也成为当前推荐系统研究的热门方法.以推荐系统中的评分预测为研究对象,通过分析指出了现有的基于图卷积神经网络的推荐模型存在2个方面的不足:图卷积层仅仅利用了1阶协同信号和未考虑用户观点的差异.为此,提出一种端到端的、基于增强图卷积神经网络的协同推荐模型.它采用一种增强的图卷积层,不仅聚合了2阶协同信号而且融合用户观点的影响,从而更合理地利用协同信号学习实体节点的嵌入表示,并通过堆叠多个图卷积层对其进行精化;最后,采用了非线性的多层感知机实现评分预测.基于5种推荐数据集上的实验结果表明:新模型的预测误差相比于几种主流的推荐模型具有明显的降低.

关键词: 协同推荐, 图卷积神经网络, 协同信号, 用户观点, 评分预测

Abstract: Graph convolutional network is a deep learning model for graph structure data, and it has become a very hot approach in the research of recommendation system due to its powerful ability in feature extraction and representation of data. This paper focuses on the rating prediction tasks in recommendation system, and points out two deficiencies of existing graph convolutional network based recommendation models by detailed analysis, including making use of first-order collaborative signals only and ignoring the differences of user opinions. For solving them, an end-to-end enhanced graph convolutional network based collaborative recommendation model is proposed. It adopts an enhanced graph convolutional layer to take full advantage of collaborative signals to learn embeddings of users and items on graph, which aggregates second-order collaborative signals and incorporates the influence of different user opinions. And it also stacks several graph convolutional layers to iteratively refine the embeddings and finally uses a nonlinear multilayer perceptron network to make rating prediction. The experiments on 5 benchmark recommendation datasets show that the proposed model achieves lower prediction errors compared with several state-of-the-art graph convolutional network based recommendation models.

Key words: collaborative recommendation, graph convolutional neural network, collaborative signal, user opinion, rating prediction