Project Title

A Graph-Aware Continual Learning Evaluation Framework

Overview

This repository contains the code implementation for the AGALE: A Graph-Aware Continual Learning Evaluation Framework. The paper introduces a framework that accommodates both single-labeled and multi-labeled nodes, addressing the limitations of previous evaluation frameworks in CGL, and the code here is intended to reproduce the experimental results presented in the paper.

Getting Started

Prerequisites

• dgl==1.0.2
• matplotlib==3.7.1
• numpy==1.23.5
• pandas==1.3.5
• pipeline==0.1
• scikit_learn==1.2.2
• scipy==1.10.1
• seaborn==0.13.1
• torch==1.12.1
• torch_geometric==2.3.0
• torch_geometric_temporal==0.54.0

Installation

1. Clone the repository:

   git clone [email protected]:Tianqi-py/AGALE.git
   cd AGALE

2. Install dependencies:

   pip install -r requirements.txt

Usage

All the baselines used in this paper are implemented in the baselines/. The backbone GNNs plugged into the baselines are in backbones/.

The folder baselines/models/ contains all the baseline models reimplemented in the two incremental settings proposed in this paper with the expandable output units for the unknown number of classes coming in the future time steps.

The name of the Python file in baselines/XX.py indicates the name of the baseline and the incremental setting in which it is implemented. For example, LwF_TaskIL is the script to run LwF in TaskIL.

Example 1: To run a baseline model in an incremental setting, run the Python file in baselines/.

Run LwF in TaskIL:

python3 LwF_TaskIL.py --backbone GCN --n_cls_per_t 2 --lr 0.01

Run LwF in ClassIL:

python3 LwF_ClassIL.py --backbone GCN --n_cls_per_t 2 --lr 0.01

Data Partition Algorithms

In this paper, we have a unified data partition algorithm for two incremental settings. The difference is how the label vectors of the nodes are handled. In the ClassIL, the nodes are allowed to have expandable label vectors, while in the TaskIL, only the classes belonging to the current task are shown. In the DataPrepare/, we provide the Python scripts to split a static graph dataset for the CGL training and testing.

Example2: prepare Yelp for CGL:

python3 Preprocessor_Yelp.py --n_cls_per_t 2 --train_ratio=0.6

where --n_cls_per_t controls how many classes will be seen at each time step. The more classes to see at each time step, the shorter the task sequence we generate from one dataset. Thus, we use two classes at each time step in this work. The --train_ratio controls the percentage of data from the given dataset for training data.

You can use the scripts as examples to prepare other graph datasets to test your model in incremental settings. Have fun!

Please cite our paper using the following bibtex code:

@article{
zhao2024agale,
title={{AGALE}: A Graph-Aware Continual Learning Evaluation Framework},
author={Tianqi Zhao and Alan Hanjalic and Megha Khosla},
journal={Transactions on Machine Learning Research},
issn={2835-8856},
year={2024},
url={https://openreview.net/forum?id=xDTKRLyaNN},
note={}
}