Multi-structure Graph Classification Method with Attention-based Pooling

This is a PyTorch implementation of MAC algorithm, which learns a graph-level representation for the entire graph. Specifically, the graph pooling operator adopts multiple strategies to evaluate the importance of nodes and update node representations through attention mechanism. Then a hierarchical architecture is designed to capture multiple different substructures of the input graph. Finally, a 2D CNN is used to generate a graph-level representation.

Requirements

• python3.7
• torch==1.7.0
• dgl==0.6.1

Note:

This code repository is built on dgl, which is a Python package built for easy implementation of graph neural network model family. Please refer here for how to install and utilize the library.

Datasets

Graph classification benchmarks are publicly available at here.

Run

To run MAC, just execute the following command for graph classification task:

python main.py

Parameter Settings

Datasets	batch_size	conv_channel1	conv_channel2	dropout_ratio	lr	pooling_ratio
PROTEINS	64	16	8	0.0	0.001	0.3
DD	64	8	4	0.7	0.001	0.6
NCI1	32	32	16	0.3	0.0005	0.4
NCI109	32	16	8	0.2	0.001	0.6
Mutagenicity	32	8	32	0.5	0.001	0.5

Note:

PROTEINS, NCI1, NCI109 and Mutagenicity are trained on NVIDIA GeForce RTX 2080 Ti GPU, DD is trained on NVIDIA Tesal V100 GPU.