When to Pre-Train Graph Neural Networks? From Data Generation Perspective!

About

This project is the implementation of paper “When to Pre-Train Graph Neural Networks? From Data Generation Perspective!" accepted by KDD2023.

Abstract

In recent years, graph pre-training has gained significant attention, focusing on acquiring transferable knowledge from unlabeled graph data to improve downstream performance. Despite these recent endeavors, the problem of negative transfer remains a major concern when utilizing graph pre-trained models to downstream tasks. Previous studies made great efforts on the issue of what to pre-train and how to pre-train by designing a variety of graph pre-training and fine-tuning strategies. However, there are cases where even the most advanced "pre-train and fine-tune" paradigms fail to yield distinct benefits. This paper introduces a generic framework W2PGNN to answer the crucial question of when to pre-train (i.e., in what situations could we take advantage of graph pre-training) before performing effortful pre-training or fine-tuning. We start from a new perspective to explore the complex generative mechanisms from the pre-training data to downstream data. In particular, W2PGNN first fits the pre-training data into graphon bases, each element of graphon basis (i.e., a graphon) identifies a fundamental transferable pattern shared by a collection of pre-training graphs. All convex combinations of graphon bases give rise to a generator space, from which graphs generated form the solution space for those downstream data that can benefit from pre-training. In this manner, the feasibility of pre-training can be quantified as the generation probability of the downstream data from any generator in the generator space. W2PGNN offers three broad applications: providing the application scope of graph pre-trained models, quantifying the feasibility of pre-training, and assistance in selecting pre-training data to enhance downstream performance. We provide a theoretically sound solution for the first application and extensive empirical justifications for the latter two applications.

Setup

The script has been tested running under Python 3.6.5, with the following packages installed (along with their dependencies):

pytorch==1.6.0

torch-geometric==1.6.3

rdkit==2020.09.1.0

networkx==2.3

scikit-learn==0.20.3

You can install the dependency packages with the following command:

pip install -r requirements.txt

File Folders

node classification: contains the code of estimating graph pre-training feasibility and pre-traing&fine-tuning data for graph classification
graph classification: contains the code of estimating graph pre-training feasibility and pre-traing&fine-tuning data for node classification
utils:contains the code of models.

Dataset

For the graph classiciation pre-training and downstream dataset, download from pre-training data, unzip it, and put it under graph_classification/data/dataset/

For the node classiciation pre-training dataset, the original datasets are stored in data.bin. And the datasets can be download through pre-training data, unzip it, and put it under node_classification/data/dataset/

Usage: How to run the code

python estimate_feasiblity.py -h

usage: estimate_feasiblity.py [-h][--pre-data][--down-data][--pre-path][--down-path][--save-path]
[--device][--seed][--method][--func][--r][--domain-num][--epoch][--learning-rate][--weight-decay]
optional arguments:
  -h, --help                show this help message and exit
  --pre-data                str, pretrain data name. 
  --down-data               str, downstream data name. 
  --pre-path                str, file path to load pre-training data.
  --down-path               str, file path to load downstream data.
  --save-path               str, file path to save graphon.
  --file-path               str, Path to save raw data.
  --device                  int, which gpu to use if any (default: 0).
  --seed                    int, set the random seed.
  --method                  str, method to estimate graphon.
  --func                    bool, whether use graphon or step function.
  --r                       str, the resolution of graphon.
  --domain-num              int, the number of domains of pre-training data
  --epochs                  int,  number of trails to fit the final graphon
  --learning-rate           float, the learning rate
  --weight-decay            float, weight decay
  --variant                 str, variant of feasibility

Demo

To run the node_classification, see example below

python node_classification/estimate_feasiblity.py --pre-datasets imdb facebook --down-data h-index

To run the graph_classification, see example below

python graph_classification/estimate_feasiblity.py --pre-data zinc_standard_agent01 --down-data bbbp

zjunet/W2PGNN