MuHca: Mixup Heterogeneous Graphs for Contrastive Learning with Data Augmentation

Contrastive learning has become a highly promising learning paradigm and demonstrated significant potential when few labels are available. The effectiveness of contrastive learning on graphs is largely dependent on the quality of positive and negative pairs, which can be improved by developing data augmentation (DA). However, the majority of the current DA methods rely on homogeneous graphs while less on heterogeneous graphs. In this paper, we present a method named MuHca, a node augmentation module for the problem of heterogeneity in DA. Concretely, MuHca separately employs nodes embedding of two views, namely meta-path and network schema, into a novel contrasting generative adversarial nets structure to implement data augmentation. By adopting the contrasting generative paradigm of GANs, MuHca can generate and optimize effective negatives. To enhance the robustness of MuHca, we exploit the potential information from the original data and extend our approaches by mixing up generated negatives with original ones. The final stage involves training a generator for edges in parallel with the modeling of edge presence among nodes, culminating in contrastive learning for heterogeneous graphs. The conducted experiments on three datasets validate the effectiveness of our proposed method, surpassing various state-of-the-art and even semi-supervised methods.