pre-training graph neural networks for computer network tasks

What is it about?

We propose a solution to the overreliance on labeled data and poor generalization of graph neural networks (GNNs) when trained in a supervised manner on computer network tasks. We exploit unlabeled, out-of-context data sources by pre-training a GNN in a self-supervised manner using link prediction. We show that a pretrained GNN needs less labeled data to learn a computer network task in a completely new setting.

Featured Image

Photo by Logan Voss on Unsplash

Photo by Logan Voss on Unsplash

Why is it important?

We are the first to propose general, self-supervised pretraining of Graph Neural Networks for computer network tasks. Since quality labeled data is extremely scarce in the computer networks field, we believe a thorough large-scale version similar to what we propose is the only way for AI to make an actual impact in the field. Our results show the promise of this direction, though it should be noted a large path to actual operational implementation is still far away.

Perspectives