基于潜在组分配及对比学习增强的符号二值图神经网络

Signed bipartite graph neural network enhanced bypotential group assignment and contrast learning

针对符号二值网络的节点异质性及三角形形式平衡理论不适用性的问题,提出一种基于潜在类分配及对比学习增强的符号二值图神经网络模型,其通过同质和异质双空间的互相补充来充分提取网络的隐式和显式信息。在同质空间,采用可学习的潜在组对节点进行分配并将节点看做多个潜在组的组合,然后通过训练来自动挖掘节点间的信息。在异质空间,对节点邻居进行有方向区分的注意力聚合,然后采用网络重建的互信息对比学习来引导聚合过程以获得表达能力更强的表示向量。在符号链接预测任务上与多种相关模型进行对比实验,实验结果显示所提出的模型在四个真实数据集上采用四种评价指标获取的16个评价结果中,12个评价结果可以取得最优值,验证了所提出模型的有效性。

To address the problems of node heterogeneity and inapplicability of triangular form balance theory in signed bipartite network modeling,this paper proposed a signed bipartite graph neural network enhanced by potential group assignment and contrast learning,which could extract the display and implicit information fully through complementing each other with homogeneous and heterogeneous spaces.In the homogeneous space,this paper treated nodes as a combination of multiple learnable potential groups,and then mined information among nodes by training automatically.In the heterogeneous space,this paper adopted the attention aggregators with directions to aggregate information of neighbors,and then used the contrast learning for network reconstruction based on mutual information to guide the aggregation process to obtain more expressive node representations.This paper performed comparative experiments with a variety of related models on the signed link prediction task.Expe-rimental results show that it can obtain optimal values for 12 of the 16 evaluation results obtained using four evaluation metrics on four real datasets,which verifies the effectiveness of the proposed model.