多视图对比增强的异质图结构学习方法

异质图神经网络作为一种异质图表示学习的方法,可以有效地抽取异质图中的复杂结构与语义信息,在节点分类和连接预测任务上取得了优异的表现,为知识图谱的表示与分析提供了有力的支撑.现有的异质图由于存在一定的噪声交互或缺失部分交互,导致异质图神经网络在节点聚合、更新时融入错误的邻域特征信息,从而影响模型的整体性能.为解决该问题,提出了多视图对比增强的异质图结构学习模型.该模型首先利用元路径保持异质图中的语义信息,并通过计算每条元路径下节点之间特征相似度生成相似度图,将其与元路径图融合,实现对图结构的优化.通过将相似度图与元路径图作为不同视图进行多视图对比,实现无监督信息的情况下优化图结构,摆脱对监督信号的依赖.最后,为解决神经网络模型在训练初期学习能力不足、生成的图结构中往往存在错误交互的问题,设计了一个渐进式的图结构融合方法.通过将元路径图和相似度图递增地加权相加,改变图结构融合过程中相似度图所占的比例,在抑制了因模型学习能力弱引入过多的错误交互的同时,达到了用相似度图中的交互抑制原有干扰交互或补全缺失交互的目的,实现了对异质图结构的优化.选择节点分类与节点聚类作为图结构学习的验证任务,在4种...

Multi-view Contrastive Enhanced Heterogeneous Graph Structure Learning

As a heterogeneous graph representation learning method, heterogeneous graph neural networks can effectively extract complex structural and semantic information from heterogeneous graphs, and have achieved excellent performance in node classification and connection prediction tasks, which provides strong support for the representation and analysis of knowledge graphs. Due to the existence of some noise interaction or missing interaction in the heterogeneous graph, the heterogeneous graph neural network incorporates erroneous neighbor features when nodes are aggregated and updated, thus affecting the overall performance of the model. In order to solve the above problems, this paper proposes a heterogeneous graph structure learning model enhanced by multi-view contrastive. Firstly, the semantic information in the heterogeneous graph is maintained by using the meta path, and the similarity graph is generated by calculating the feature similarity between the nodes under each meta-path, which is fused with the meta-path graph to optimize the graph structure. By comparing the similarity graph and meta-path graph as different views, the graph structure is optimized without the supervision information, and the dependence on the supervision signal is eliminated. Finally, in order to solve the problem that the learning ability of neural network model is insufficient at the initial stage of training and there are often error interactions in the generated graph structure, this paper designs a progressive graph structure fusion method. Through incremental weighted addition of meta-path graph and similarity graph, we change the weight of similarity graph in the fusion process, it not only prevents erroneous interactions from being introduced in the initial stage of training, but also achieves the purpose of using the interaction in similarity graph to suppress interference interaction or complete missing interaction, thus the structure of heterogeneous graph is optimized. We select node classification and node clustering as the verification tasks of graph structure learning. The experimental results on four real heterogeneous graph datasets prove that the heterogeneous graph structure learning method proposed in this paper is feasible and effective. Compared with the optimal comparison model, the performance of our model has been significantly improved under two evaluation metrics.