计算机科学与探索 ›› 2023, Vol. 17 ›› Issue (2): 453-466.DOI: 10.3778/j.issn.1673-9418.2106010

• 人工智能·模式识别 • 上一篇    下一篇

自适应融合邻域聚合和邻域交互的图卷积网络

富坤,禚佳明,郭云朋,李佳宁,刘琪   

  1. 河北工业大学 人工智能与数据科学学院,天津 300401
  • 出版日期:2023-02-01 发布日期:2023-02-01

Graph Convolutional Network with Adaptive Fusion of Neighborhood Aggregation and Interaction

FU Kun, ZHUO Jiaming, GUO Yunpeng, LI Jianing, LIU Qi   

  1. School of Artificial Intelligence, Hebei University of Technology, Tianjin 300401, China
  • Online:2023-02-01 Published:2023-02-01

摘要: 图表示学习是在保持图中节点性质不变的前提下,获取节点的低维表示向量,为下游任务提供有效的数据支持。现有图表示学习算法大多关注于聚合邻域特征,对于挖掘其他非线性信息关注不足。针对这一问题,提出了自适应融合邻域聚合和邻域交互的图卷积网络AFAI-GCN。首先,采用双通道图卷积网络建模邻域聚合,并利用生成的嵌入表示计算邻域交互项来补充算法学习的信息;然后,结合注意力机制构建自适应融合模块,增加对重要信息的关注,提高融合信息项的任务相关性;最后,通过信息一致性约束和差异性约束增强节点特征一致性和嵌入表示差异性。在三个公共引文数据集上进行了节点分类和可视化任务,结果显示,AFAI-GCN与图卷积网络(GCN)、邻域聚合和交互图卷积网络(AIR-GCN)等算法相比,在Cora、Citeseer、Pubmed数据集上分类准确率分别提高了1.0~1.6个百分点、1.1~2.4个百分点和0.3~0.9个百分点;在可视化任务中团簇内聚合程度更高,不同的团簇边界更清晰;算法学习过程中收敛速度更快,准确率曲线更平滑。实验结果表明该框架较好地提升了基准算法的性能。

关键词: 图表示学习, 图卷积神经网络(GCNN), 注意力机制, 节点分类

Abstract: Graph representation learning technology aims to learn the low dimensional representation vectors for nodes while maintaining the properties of graphs and provide materials for downstream tasks. Most existing algorithms mainly focus on aggregating neighborhood features, lack the ability to capture non-linear information. To solve this issue, this paper proposes a graph convolutional network with adaptive fusion of neighborhood aggregation and interaction (AFAI-GCN). Firstly, a two-channel graph convolutional network is constructed to model the neighborhood aggregation, and the representations generated by modeling are used to calculate the neighborhood interaction items to enhance the learning capability of algorithm. Secondly, the attention mechanism is adopted in the adaptive fusion module to increase the attention on significant information and improve the task correlation of the fused information items. Finally, an information consistency constraint and a difference constraint are applied to enhancing node feature consistency and embedded representation difference. The node classification and visualization tasks are performed on three public citation datasets. Experiments results show that, compared with graph convolutional network (GCN), neighborhood aggregation and interaction graph convolutional network (AIR-GCN) and other algorithms, the classification accuracy of AFAI-GCN is increased by 1.0 to 1.6, 1.1 to 2.4, 0.3 to 0.9 percentage points on Cora, Citeseer and Pubmed datasets, respectively. In the visualization task, the degree of aggregation within clusters is higher and the boundaries of different clusters are clearer. Moreover, the convergence speed of AFAI-GCN is faster and the accuracy curve is smoother in the learning process. All these results indicate that AFAI-GCN is more advanced than the benchmark algorithms.

Key words: graph representation learning, graph convolutional neural network (GCNN), attention mechanism, node classification