Article Source

• Position aware Graph Neural Networks

Position aware Graph Neural Networks

• Jure Leskovec
• Computer Science, PhD
• Course web site: http://web.stanford.edu/class/cs224w/
• To get the latest news on Stanford’s upcoming professional programs in Artificial Intelligence, visit: http://learn.stanford.edu/AI.html

Abstract

We introduce the idea of Position-aware for graphs. To start, we define position-aware tasks, where we would like to classify nodes based on their positions in the graph. We demonstrate that certain position-aware tasks will always cause GNNs to fail. Our solution is the Position-aware Graph Neural Networks (P-GNN). The key idea of P-GNN is to introduce randomly selected anchors node, where we will embed all the nodes by computing their shortest path distances to these anchor nodes. To save the number of anchors needed, we further generalize the notion of anchor to anchor-sets, where each anchor-set contains a varied number of nodes.

You can find more details on the P-GNN paper. https://arxiv.org/abs/1906.04817