摘要
针对大多图对比学习方法中对输入图进行随机增强和损失函数中忽略考虑图的同质性假设的问题,提出基于邻居监督和自适应增强的图对比学习框架。该框架利用输入图中节点特征向量的中心性进行自适应增强生成2个视图,避免随机增强对重要的节点和边进行删除。利用针对图结构数据的邻居监督图对比损失函数指导框架学习,采用网络拓扑结构作为监督信号定义对比学习中的正负样本,允许每个锚点有多个正样本。所提框架在3个引文数据集上进行节点分类实验,实验表明,其在分类准确性方面优于很多基线方法。
In response to the issues of random graph augmentation and the omission of considering graph homogeneity in the loss function in most graph contrastive learning methods,a graph contrastive learning framework based on neighbor supervision and adaptive enhancement is proposed.The framework utilizes the centrality of node feature vectors in the input graph for adaptive enhancement,generating two views that prevent the deletion of important nodes and edges caused by random augmentation.The framework learns are guided by the neighbor supervised graph contrastive loss function tailored for graph-structured data,using network topology as a supervisory signal to define positive and negative samples in contrastive learning,allowing multiple positive samples for each anchor point.The proposed framework is evaluated through node classification experiments on three citation datasets,and the experimental results demonstrate its superior performance in terms of classification accuracy compared to several baseline methods.
作者
李浩然
杨艳
钟颖莉
LI Haoran;YANG Yan;ZHONG Yingli(School of Computer Science and Technology,Heilongjiang University,Harbin 150080,China)
出处
《黑龙江大学工程学报(中英俄文)》
2024年第3期36-44,共9页
Journal of Engineering of Heilongjiang University
基金
黑龙江省自然科学基金项目(LH2022F045)
关键词
对比学习
图神经网络
邻居监督
节点分类
contrastive learning
graph neural network
neighbor supervision
node classification
