DSLR: Diversity Enhancement and Structure Learning for Rehearsal-based Graph Continual Learning

Introduction

Implementation of DSLR: Diversity Enhancement and Structure Learning for Rehearsal-based Graph Continual Learning.

Continual learning method with replay approach on graph structured data. The structure learning framework and experience replay approach were used to prevent catastrophic forgetting.

Basics

1. The main train/test code is in train.py
2. If you want to see the continual GNN model in PyTorch Geometric MessagePassing grammar, refer to model.py
3. If you want to see the replay buffer selection stage, refer to replay.py
4. If you want to see hyperparameter settings, refer to train.py

Run DSLR

1. Cora
python train.py --dataset cora --replay CD --structure yes --classes_per_task 2 --memory_size 100

2. Amazon Computer
python train.py --dataset amazoncobuy --replay CD --structure yes --classes_per_task 2 --memory_size 200

3. OGB-arxiv
python train.py --dataset ogb_arxiv --replay CD --structure yes --classes_per_task 3 --memory_size 3000