GSVAELP: integrating graphSAGE and variational autoencoder for link prediction

Link prediction (LP) plays a crucial role in network science, which forecasts potential connections or relationships between nodes or entities within the network. Link prediction has found many applications, such as suggesting new connections to users in social networks, personalized recommendations in e-commerce, predicting new routes in transportation networks, and many others. This paper introduces a deep learning-based link prediction model in social networks by leveraging graphSAGE (graph sample and aggregation) and Variational Autoencoders (VAE). The proposed work starts by utilizing a graphSAGE to generate node embeddings of the input network by sampling and aggregating information from neighborhood nodes. The generated embeddings are sufficiently expressive to capture the local and global network structure. Further, we adopt VAE to learn a latent space representation of the graphSAGE embeddings. The VAE helps to refine the node embeddings and learn a meaningful latent space representation of the input data, which can be useful for the downstream link prediction task. The encoder’s output (latent space) can capture important features that aid link prediction. Finally, we train a logistic regression classifier using the latent representations from the VAE as features to predict the upcoming links in the network. The necessary hyperparameter studies are performed to obtain the optimal values of the various model parameters. The experiments and simulations conducted on eight different real-world network datasets illustrate the effectiveness of the proposed link prediction model. Additionally, we evaluate the average performance of each comparative link prediction method across all datasets to assess the efficacy of the proposed approach.