Graph convolutional neural networks (GCNs) have revolutionized the field of representation learning on graph data with non-Euclidean properties. As one of its variants, Graph attention networks (GATs) leverage masked self-attentional layers to specify different weights when aggregating node features over the neighbourhoods of graphs. GATs have achieved the state-of-the-arts results across many benchmarking datasets for the task of node classification. However, this method is insufficient in learning graph-level representations for graph classification, which is another important graph learning task. We propose a novel graph pooling method (namely PagePool) to extend GATs to perform graph classification instead of node classification. This method leverages both PageRank message passing algorithm and the attention coefficients of GATs to propagate and calculate the feature-aware node importance estimates (namely attentional PageRank). The attentional PageRank (attPR) values can then be used to select nodes from graphs to get graph-level representations for graph classification.
Copyright is held by the author(s).
This thesis may be printed or downloaded for non-commercial research and scholarly purposes.
Supervisor or Senior Supervisor
Thesis advisor: Ester, Martin
Member of collection