On an improvement of Graph Convolutional Network in semi-supervised learning

Nowadays, we can easily face with many real-world datasets having the form of graphs, such as social networks, knowledge-based graphs, protein-interaction networks, the World Wide Web, etc., while there was still lack of in-depth research on the generalization of neural network models to such structured datasets. Hopefully, in the last few years, a plenty of encouraging results devoted to generate neural networks to work on arbitrarily structured graphs. Among them, it can be seen an important scalable approach of Kipf and Welling in 2016, which employs an efficient variant of convolutional neural networks for semi-supervised learning on graph-structured data. The experimental results applied for citation networks and on a knowledge graph dataset show that their proposed approach can outperform all related methods by a significant margin. In this paper, we present an adaptive approach for semi-supervised learning on graph-structured data that is also based on an efficient variant of convolutional neural networks which improves the pilot work of Kipf and Welling. In a number of separated experiments on citation networks datasets (Citeseer, Cora, Pubmed), we demonstrate that our approach is able to outperform the baseline graph convolutional network (GCN).