Boosting Performance of Graph Convolutional Networks via Generating Pseudolabels and Feature Interaction

The task of node classification based on graph convolutional networks (GCNs) is widely applied in many social media and e-commerce websites. In this context, rich node labels undoubtedly play a very important role. However, existing methods fail to generate high-quality pseudolabels that serve as a potential solution to alleviate label scarcity on graphs and reduce time-consuming annotation issues. Moreover, it is an interesting problem to boost the performance of GCNs with feature interaction and propagation. To tackle these limitations, we propose a new pseudolabels generation and feature interaction propagation based GCN, named PF-GCN. Specifically, the proposed PF-GCN first generates high-quality pseudolabels to enlarge the input set of node labels. Meanwhile, it fully exploits the information of feature interactions to enhance our PF-GCN model. We then present a straightforward yet highly effective cross-layer approach which is equipped with feature transformation to encode feature interaction at each layer. Finally, we propagate feature interactions through both topology space and feature space, and then employ an adaptive attention mechanism to learn rich node embeddings. Experimental results demonstrate that the proposed PF-GCN outperforms several state-of-the-art methods in the extensive evaluations of node classification. We have released the source codes of PF-GCN for public usage and evaluation at https://github.com/ZZY-GraphMiningLab/PF-GCN.