Multi-Head Variational Graph Autoencoder Framework for Link Prediction on Citation Graphs
Abstract
Graph neural networks (GNNs) are powerful tools for link prediction in diverse applications such as recommendation systems, drug discovery, knowledge graph completion, and more. Link prediction estimates the likelihood of node connections, enabling various downstream tasks and providing insights into network structures. However, generating highly representative node embeddings in a semi-supervised setting with limited labelled data is challenging. GNNs, utilizing neural message passing, effectively leverage local and global graph information to address this issue. Our research aims to enhance node embeddings for link prediction by developing novel techniques that capture intricate graph patterns, incorporating both; topological and semantic information. We leverage the expressive power of GNNs to overcome the limitations of traditional methods and improve link prediction accuracy. The proposed model was tested for three benchmark datasets, namely Cora, CiteSeer and PubMed, for the link prediction tasks on citation graphs. It is observed that the proposed model outperforms the existing models for this task by achieving area under curve (AUC) and average precision (AP) scores of 99.1%, 99.2% for Cora, 98.5%, 98.3% for Citeseer and 99.1%, 98.9% for the PubMed dataset respectively. The average improvement of 2% and 3% is observed in AUC and AP scores, respectively. The proposed model outperforms the existing methods and can be extended further for other datasets.