Fachhochschule Nordwestschweiz FHNW

Link Prediction in Graphs with Graph Neural Networks

Summary

Applying Transfer Learning to Enhance Link Prediction with GNNs.

Keywords

Transfer Learning, Graph Theory, Graph Neural Networks, Deep Learning

Objective

Our goal is to demonstrate the effectiveness of transfer learning in predicting connections (links) between data points in networks. By using pre-trained Graph Neural Networks, we explore whether and how existing models can be adapted to new, more specific datasets to improve the accuracy and efficiency of predictions.

Background

Analyszing and predicting connections in large datasets is a challenging task that is important in many scientific and commercial applications. Traditional models often reach their limits when it comes to effectively modeling the complex relationships in graph data. Transfer learning offers a promising approach to utilise models already trained on large and general datasets and optimise them for more specific tasks.

Results

Our results demonstrate a substantial reduction in training times due to the application of transfer learning with Graph Neural Networks. While the enhancement in prediction accuracy was not significant, the efficiency gains from using pre-trained models were notable. This approach allowed us to leverage complex pre-existing models on new datasets effectively, illustrating that even without major accuracy improvements, the time savings alone offer a considerable advantage for practical applications of GNNs.