基于正负样本选择的异质图对比学习  

Heterogeneous graph contrastive learning based on positive and negative sample selection

作　　者：宋笑笑 刘勇[1] 金虎[1] SONG Xiaoxiao;LIU Yong;JIN Hu(School of Computer Science and Technology,Heilongjiang University,Harbin 150080,China)

机构地区：[1]黑龙江大学计算机科学与技术学院,哈尔滨150080

出　　处：《黑龙江大学工程学报(中英俄文)》2024年第3期45-53,共9页Journal of Engineering of Heilongjiang University

基　　金：黑龙江省自然科学基金项目(LH2020F04)

摘　　要：对比学习应用于异质图能有效避免训练过程中对标注数据的依赖性,无需依赖大量带标签样本的前提下,依旧能够提取异质图中复杂的结构和丰富的语义,在节点分类和节点聚类等任务上表现出优越的性能。现有的异质图对比模型直接基于元路径构造对比视图,用来捕获对象之间的语义以及序列关系。此类方案忽略了网络的局部结构,目标节点没有完全融合所有一阶邻居节点的信息。同时,局限于单一的正样本,无法区分负样本的正确性,使模型出现梯度消失或者学习内容不全。为解决上述问题,提出基于正负样本选择的异质图对比学习模型。该模型先构造了网络模式视图和元路径视图,用来保持异质图的局部结构和高阶结构;以网络模式视图中的节点为锚点,在元路径视图中构造正负样本,依据样本的相似性来学习负样本的权重;实验证明提出的模型优于当前的基线方法。Contrastive learning applied to heterogeneous graph can effectively avoid dependency on labeled data during the training process.Even without relying on a large number of labeled samples,it can still extract complex structures and rich semantics from heterogeneous graph,demonstrating superior performance in tasks such as node classification and node clustering.Existing heterogeneous graph contrastive models directly construct contrastive views based on meta-paths to capture the semantics and sequential relationships between objects.However,such approaches often ignore the local structure of the network,and the target node does not fully integrate information from all first-order neighbor nodes.Moreover,limited to a single positive sample,they can’t distinguish the correctness of negative samples,leading to gradient vanishing or incomplete learning content.To address these issues,this paper proposes a heterogeneous graph contrastive learning model based on the selection of positive and negative samples.Firstly,the model constructs network pattern views and meta-path views to preserve the local structure and higher-order structure of the heterogeneous graph.Secondly,using nodes in the network pattern view as anchor nodes,positive and negative samples are constructed in the meta-path view,and the weights of negative samples are learned based on the similarity of the samples.Finally,extensive experiments are conducted to demonstrate that the proposed model outperforms the current state-of-the-art methods.

关 键 词：异质图 对比学习 难负样本 图神经网络 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]