Expanding Semantic Knowledge for Zero-Shot Graph Embedding

Zero-shot graph embedding is a major challenge for supervised learning. Although recent method RECT has shown promising performance, its working mechanisms are not clear and still needs lots of training data. In this paper, we give deep insights into RECT, address fundamental limits. We show that core part GNN prototypical model in which class prototype described by mean feature vector. As such, maps nodes from the raw-input space an intermediate-level semantic connects features to both seen unseen classes. This mechanism makes work well on classes, however also reduces discrimination. To realize full potentials, propose two label expansion strategies. Specifically, besides expanding labeled node set can expand Experiments real-world datasets validate superiority our methods.