A study on hybrid graph neural network-based approach for link prediction

Purpose The purpose of the presented study is to explore graph and analyse the performance of hybrid graph neural networks for link prediction tasks in a graph. Graphs are a collection of nodes and edges that form a kind of network and represent relationships. Link prediction in graphs plays a vital role in various applications like recommender system, connectivity prediction, molecular bonding etc. In recent times, graph neural networks have become a popular approach for link prediction. Design/methodology/approach In this work, hybrid graph neural networks are developed by combining various standard graph neural networks. Graph Convolution Network (GCN), Graph Attention Network (GAT), Graph Isomorphism Network (GIN), Graph Sample and Aggregate (GraphSAGE) and Edge Convolution Network (EdgeConv) models are individually executed with three embedding approaches named Standard, Node2Vec and DeepWalk for link prediction task. Hybrid graph neural network is designed using stacking and ensembling approaches on the previously mentioned five models as layers of proposed hybrid model. Experiments are performed on three benchmark datasets Cora, CiteSeer and PubMed. Findings A comparative study is also carried out by comparing the individual models with the hybrid models developed using stacking and ensemble approaches. Area Under the curve (AUC) score is the primary metrics used to measure the performance of various individual and hybrid graph neural network models. Originality/value It is observed from the experiments that hybrid graph neural networks with an ensemble approach have achieved better AUC than the stacking approach as well as individual graph neural network models for the link prediction task.