【摘要】 无监督图表示学习因无需人工标注，成为当前研究的热点方向.图对比学习通过正负样本对构建，实现结构与语义一致性的多视角建模，成为主流方法之一.针对现有异质图对比学习依赖人工元路径设计、难以灵活适应多元结构与语义的局限，本文通过边分类器动态预测边的同质或异质属性，并据此构建同质视图与异质视图，并结合双视图交叉对比学习实现结构与语义的统一建模，提升表示的判别性与鲁棒性.在5个同质图和4个异质图数据集上的实验表明，带有边分类器的图对比学习(graph contrastive learning with edge classifier,ECGCL)在同质图上性能与主流基线方法持平;在异质图上，节点分类准确率较同质方法最高提升26.3%，较异质基线方法最高提升4.8%，验证了其有效性与泛化能力.更多还原

【Abstract】 Unsupervised graph representation learning has become a research hotspot because it does not require manual annotation.Graph contrastive learning has become one of the mainstream methods by constructing positive and negative sample pairs to achieve multi-view modeling of structural and semantic consistency.However,existing heterogeneous graph contrastive learning methods rely on manual meta-path design and struggle to flexibly adapt to diverse structural and semantic limitations.The homogeneous or heterogeneous properties of edges were dynamically predicted through an edge classifier,based on which a homogeneous view and a heterogeneous view were constructed.By integrating cross-view contrastive learning between these dual views,unified modeling of structural and semantic information was achieved,thereby enhancing the discriminativeness and robustness of the learned representations.Experiments on five homogeneous and four heterogeneous graph datasets show that ECGCL(graph contrastive learning with edge classifier) achieves performance comparable to mainstream baselines on homogeneous graphs,and on heterogeneous graphs,it improves node classification accuracy by up to 26.3% over homogeneous methods and up to 4.8% over heterogeneous baselines,demonstrating its effectiveness and generalization ability.更多还原