基于图聚类结果的半监督节点分类方法

Semi-Supervised Node Classification Method Based on Graph Clustering Results

目前在处理现实世界中知识图谱、引文网络以及社交网络等复杂的图结构数据分类问题上,图卷积神经网络被认为是最有效的半监督方法之一,但存在其学习性能会被严重有限的标记数据影响的问题。本研究针对这一问题提出了一种以图聚类结果作为指导的节点分类方法。具体来说,引入数据增强模块减少了图结构信息中的噪声,设计了一种面向聚类的图嵌入模型作为属性图聚类网络,并根据聚类结果预测出节点的伪标签。同时,为了提升分类任务的性能,筛选出高置信度伪标签来指导图节点分类任务,并设计了一种相似度损失来提高标记节点和未标记节点之间的特征相似度。通过在基准数据集上大量的实验结果表明,与现有的方法相比,该方法可以克服标签数量限制,在图节点分类任务上表现出优越的性能。Currently, graph convolutional neural networks are considered to be one of the most effective semi-supervised methods in dealing with the classification of complex graph-structured data in real-world, such as knowledge graphs, citation networks, and social networks, but there is still the problem that their learning performance can be affected by severely limited labeled data. In this study, a node classification method using graph clustering results as a guide is proposed to address this problem. Specifically, a data enhancement module is introduced to reduce the noise in the graph structure information, a clustering-oriented graph embedding model is designed as an attribute graph clustering network, and the pseudo-labels of the nodes are predicted based on the clustering results. Meanwhile, to improve the performance of the classification task, high-confidence pseudo-labels are screened to guide the graph node classification task, and a similarity loss is designed to improve the feature similarity between labeled and unlabeled nodes. The results of extensive experiments on the benchmark dataset show that the method can overcome the label number limitation and have superior performance on the graph node classification task compared with existing methods.