CODE-SUBMIT/Graph_Neighborhood_1

First and Complementary Neighborhood Combination of Adjacency Matrix for Graph Learning

Mutual Information Maximization in Graph Neural Networks

Overview

This repository contains the implementation of paper titled 'Mutual Information Maximization in Graph Neural Networks', which was accepted by IJCNN 2020. In the paper, we extend the graph neural networks frameworks by exploring the aggregation and iteration scheme in the methodology of mutual information. We propose a new approach of enlarging the normal neighborhood in the aggregation of graph neural networks, which aims at maximizing mutual information. The proposed approach improves the performance of the following graph models:

• GCN from Xu et al.: Representation learning on graphs: Methods and applications (2017)

• GIN from Xu et al.: Representation learning on graphs: Methods and applications (ICLR-2019)

• LDS-GNN from Luca et al.: Learning Discrete Structures for Graph Neural Networks (ICML-2019)

• GMNN from Luca et al.: Graph Markov Neural Networks (ICML-2019)

• PWL from Bastian Rieck et al.: A Persistent Weisfeiler–Lehman Procedure for Graph Classification (ICML-2019)

• GRAPH_Unet Gao et al.: Graph_Unet (ICML-2019)

• Graphite Grover et al.: Graphite: Iterative Generative Modeling of Graphs (ICML-2019)

• VGAE Max Welling et al.: VGAE:Variational graph auto-encoders (ICML-2019)

• MGCNK Max Welling et al.: Semi-Supervised Classification with Graph Convolutional Networks (ICLR-2017)

• CHEBNET Boris Knyazev et al.: Spectral Multigraph Networks for Discovering and Fusing Relationships in Molecules (NipsW-2018)

Different models are in separated folders.

Experimental results

Experiment No. 1:

Supervised graph classification in comparison with GCN and GIN on 7 datasets.

Experiment No. 2:

Supervised graph classification in comparison with KNN-LDS on 6 datasets.

Experiment No. 3:

Supervised graph classification in comparison with P-WL and its variants on 2 datasets.

Experiment No. 4:

Semi-supervised graph classification in comparison with GMNN on 3 datasets.

Experiment No. 5:

Graph link prediction.

Experiment No. 6:

Edge generation and graph classification.

Experiment No. 7:

Graph classification with node attribute.

Experiment No. 8:

Graph classification in comparison with GCN, MGCN and MGCNK.

Experiment No. 9:

Supervised graph classification in comparison with two transformation forms and two baseline models.

Experiment No. 10:

Supervised graph classification for three datasets in comparison with Mixhop and (s)gmnn.

Experiment No. 11:

Supervised graph classification in comparison with the state-of-the-art models on 13 datasets.