Optimal Transport Guided Node Classification in Cross Networks

Inspired by the huge powerful representation ability of graph neural networks (GNNs), GNNs begin to solve node classification in cross networks. However, current GNNs are designed for single graph, and overlook the distribution shift in cross network which degrades the performance on test nodes. In this paper, we firstly analyze deep cross-network embeddings based on GNNs and illustrate that in cross-network node classification training and test embedding spaces have different distributions. According to domain adaptation and our analysis, we propose a novel framework and loss function, which introduces optimal transport to GNNs and generates transferable embeddings. Extensive experiments in citation networks verify the effectiveness of our method OT-DCNE.