Bet-GAT: An Efficient Centrality-Based Graph Attention Model for Semi-Supervised Node Classification

Graph Neural Networks (GNNs) have witnessed great advancement in the field of neural networks for processing graph datasets. Graph Convolutional Networks (GCNs) have outperformed current models/algorithms in accomplishing tasks such as semi-supervised node classification, link prediction, and graph classification. GCNs perform well even with a very small training dataset. The GCN framework has evolved to Graph Attention Model (GAT), GraphSAGE, and other hybrid frameworks. In this paper, we effectively usd the network centrality approach to select nodes from the training set (instead of a traditional random selection), which is fed into GCN (and GAT) to perform semi-supervised node classification tasks. This allows us to take advantage of the best positional nodes in the network. Based on empirical analysis, we choose the betweenness centrality measure for selecting the training nodes. We also mathematically justify why our proposed technique offers better training. This novel training technique is used to analyze the performance of GCN and GAT models on five benchmark networks-Cora, Citeseer, PubMed, Wiki-CS, and Amazon Computers. In GAT implementations, we obtain improved classification accuracy compared to the other state-of-the-art GCN-based methods. Moreover, to the best of our knowledge, the results obtained for Citeseer, Wiki- CS, and Amazon Computer datasets are the best compared to all the existing node classification methods.

Automated summary

Graph Neural Networks (GNNs) have made significant progress in processing graph datasets, and Graph Convolutional Networks (GCNs) have outperformed other models in tasks such as node classification, link prediction, and graph classification. A novel training technique based on network centrality is proposed in this paper, which improves the performance of GCN and GAT models. Empirical analysis shows that the proposed technique achieves better classification accuracy compared to existing methods, and it sets new records on benchmark datasets.