PiNet: Attention Pooling for Graph Classification

TL;DR

PiNet introduces an attention-based pooling mechanism for graph neural networks, enhancing graph classification accuracy and efficiency, especially in distinguishing isomorphic graphs, with competitive results on chemo-informatics datasets.

Contribution

PiNet presents a novel differentiable attention pooling method that improves graph classification performance and sample efficiency over existing GNN approaches.

Findings

High sample efficiency in graph classification.

Superior performance in distinguishing isomorphic graphs.

Competitive results on chemo-informatics datasets.

Abstract

We propose PiNet, a generalised differentiable attention-based pooling mechanism for utilising graph convolution operations for graph level classification. We demonstrate high sample efficiency and superior performance over other graph neural networks in distinguishing isomorphic graph classes, as well as competitive results with state of the art methods on standard chemo-informatics datasets.