RGE: A Repulsive Graph Rectification for Node Classification via Influence

Introduction

Official JAX implementation of ICML 2023 paper "RGE: A Repulsive Graph Rectification for Node Classification"

This work scrutinizes the trend that there exists an influence estimation error of an edge group in graph influence function and this error might decrease the peroformance in graph rectification. However, for a pair of distant edges (where each edge affects the representation of different train nodes), we observe and theoretically show the estimation error is zero. Thus, we propose RGE, which eliminates distant edges at each iteration in graph rectification, and demonstrate the effectiveness of RGE on various graphs.

Commands

• Pretraining SGC

  CUDA_VISIBLE_DEVICES=0 python -m pretrain.main \
  --dataset Cora \
  --data_dir ./data \
  --lr 0.2 \
  --weight_decay 8.5e-6
  

• Identifying the opponent edges (lr and weight decay should be the same with HP used in pretraining SGC)

  CUDA_VISIBLE_DEVICES=0 python -m rge.main_retrain \
  --dataset Cora \
  --data_dir ./data \
  --lr 0.2 \
  --weight_decay 8.5e-6
  

Hyper-Parameters (HP) for pretraining SGC

• Homophilous graphs

HP (SGC)	Cora	CiteSeer	PubMed	Photo	Computers
learning rate	0.2	0.5	0.2	0.2	0.2
weight decay	8.5e-6	8.5e-6	8.5e-6	5e-7	1e-7

• Heterophilous graphs

HP (SGC)	Cornell	Wisconsin	Texas	Actor	Squirrel
learning rate	0.2	0.2	0.2	0.2	0.2
weight decay	2e-6	3e-5	4e-5	1e-5	1e-7

Performance of pretrainted SGC

• Homophilous graphs

Method (Acc.)	Cora	CiteSeer	PubMed	Photo	Computers
SGC	81.00	71.90	78.90	90.22	86.65
RGE	84.50	73.75	82.80	91.64	88.85

• Heterophilous graphs

Method (Acc.)	Cornell	Wisconsin	Texas	Actor	Squirrel
SGC	54.32	64.90	63.24	31.21	40.54
RGE	56.32	68.84	63.24	31.64	40.93

Dependencies

This code has been tested with

• Python == 3.8.10
• JAX == 0.3.17
• Haiku == 0.0.9.dev
• Jraph == 0.0.6.dev0
• Pytorch == 1.13.0+cu117
• Pytorch Geometric == 2.1.0