Zero-shot Heterogeneous Graph Embedding via Aggregating Metapath Semantically

Many real-world graphs exhibit inherent heterogeneity, involving a diverse range of node types and relation types. Existing graph neural networks are primarily designed for balanced datasets, where each category has at least one labeled node. However, real-world datasets often present complete imbalances, with certain node types lacking labeled examples. This poses a significant challenge for traditional graph embedding models, leading to diminished performance across various downstream tasks. In this study, we introduce a novel zero-shot heterogeneous graph embedding method, named Zero-shot Semantically Metapath-Aggregating graph neural network (ZSMA), which incorporates a semantic module into the conventional heterogeneous graph embedding network. This augmentation facilitates the extraction of shared features among nodes of the same type in the original graph. By mapping nodes of the same type to a lower-dimensional latent vector space through the extraction of semantic features, ZSMA enhances the performance of the graph network in zero-shot (completely-imbalanced) scenarios. We further validate the effectiveness of the proposed approach through experiments on real-world datasets.