• Stars
    star
    196
  • Rank 198,553 (Top 4 %)
  • Language
    Python
  • License
    MIT License
  • Created about 2 years ago
  • Updated 9 months ago

Reviews

There are no reviews yet. Be the first to send feedback to the community and the maintainers!

Repository Details

Code for NeurIPS 2022 paper "Knowledge Distillation Improves Graph Structure Augmentation for Graph Neural Networks"

Knowledge Distillation for Graph Augmentation (KDGA)

This is a PyTorch implementation of the Knowledge Distillation Improves Graph Augmentation (KDGA), and the code includes the following modules:

  • Dataset Loader (Cora, Citeseer, Texas, Cornell, Wisconsin, Actor, Chameleon, and Squirrel)

  • GCN Classifier for implementing $p(Y|A,X)$, and Graph Augmentation Module for implementing $p(\widehat{A}|A,X)$

  • Training paradigm for pre-training and fine-tuning on eight real-world datasets

  • Visualization and evaluation metrics

Introduction

Graph (structure) augmentation aims to perturb the graph structure through heuristic or probabilistic rules, enabling the nodes to capture richer contextual information and thus improving generalization performance. While there have been a few graph structure augmentation methods proposed recently, none of them are aware of a potential \textit{negative augmentation} problem, which may be caused by overly severe distribution shifts between the original and augmented graphs. In this paper, we take an important graph property, namely graph homophily, to analyze the distribution shifts between the two graphs and thus measure the severity of an augmentation algorithm suffering from negative augmentation. To tackle this problem, we propose a novel Knowledge Distillation for Graph Augmentation (KDGA) framework, which helps to reduce the potential negative effects of distribution shifts, i.e., negative augmentation problem. Specifically, KDGA extracts the knowledge of any GNN teacher model trained on the augmented graphs and injects it into a partially parameter-shared student model that is tested on the original graph. As a simple but efficient framework, KDGA is applicable to a variety of existing graph augmentation methods and can significantly improve the performance of various GNN architectures. For three popular graph augmentation methods, the experimental results show that the learned student models outperform their vanilla implementations by an average accuracy of 4.6% (GAUG), 4.2% (MH-Aug), and 4.6% (GraphAug) on eight graph datasets.

Dependenciesc

  • numpy==1.19.2
  • scipy==1.3.1
  • torch==1.6.0
  • pyro==1.3.0

Overview

  • main.py

    • pretrain_Augmentor() -- Pretrain Graph Augmentation Module - GraphAug
    • pretrain_Classifier() -- Pretrain GNN Classifier
    • main() -- Train the model for node classification task on eight real-world datasets
  • model.py

    • GCNLayer() -- GCN Layer
    • GCN_Classifier() -- GCN Classifier for implementing the function $p(Y|A,X)$
    • Augmentor() -- Graph Augmentation Module - GraphAug for implementing the function $p(\widehat{A}|A,X)$
    • com_distillation_loss() -- Calculate the KL-divergence Loss for knowledge distillation
  • dataset.py

    • load_data() -- Load Cora, Citeseer, Texas, Cornell, Wisconsin, Actor, Chameleon, and Squirrel datasets
  • utils.py

    • evaluation() -- Calculate classification accuracy

Running the code

  1. Install the required dependency packages

  2. To get the results on a specific dataset, please run with proper hyperparameters:

python main.py --dataset data_name --loss_mode mode

where the data_name is one of the eight datasets (Cora, Citeseer, Texas, Cornell, Wisconsin, Actor, Chameleon, and Squirrel), and loss_mode denotes different experimental settings (-1: default optimal hyperparameters obtained by NNI; 0: w/ parameter-shared DKGA; 1: w/ parameter-independent DKGA; 2: vanilla GraphAug; 3: vanilla GCN). Use the default optimal hyperparameters on the Citeseer dataset as an example:

python main.py --dataset citeseer --loss_mode -1

Citation

If you find this project useful for your research, please use the following BibTeX entry.

@article{wu2022knowledge,
  title={Knowledge Distillation Improves Graph Structure Augmentation for Graph Neural Networks},
  author={Wu, Lirong and Lin, Haitao and Huang, Yufei and  and Li, Stan Z.},
  journal={Advances in Neural Information Processing Systems},
  year={2022}
}

More Repositories

1

awesome-graph-self-supervised-learning

Code for TKDE paper "Self-supervised learning on graphs: Contrastive, generative, or predictive"
1,345
star
2

awesome-protein-representation-learning

Awesome Protein Representation Learning
517
star
3

MAPE-PPI

Code for ICLR 2024 (Spotlight) paper "MAPE-PPI: Towards Effective and Efficient Protein-Protein Interaction Prediction via Microenvironment-Aware Protein Embedding"
Python
112
star
4

GCML

Code for WACV 2022 paper "Generalized Clustering and Multi-Manifold Learning with Geometric Structure Preservation"
Python
18
star
5

GraphMixup

Code for ECML-PKDD 2022 paper "GraphMixup: Improving Class-Imbalanced Node Classification by Reinforcement Mixup and Self-supervised Context Prediction"
Python
16
star
6

FF-G2M

Code for AAAI 2023 (Oral) paper "Extracting Low-/High- Frequency Knowledge from Graph Neural Networks and Injecting it into MLPs: An Effective GNN-to-MLP Distillation Framework"
Python
16
star
7

KRD

Code for ICML 2023 paper "Quantifying the Knowledge in GNNs for Reliable Distillation into MLPs"
Python
11
star
8

RFA-GNN

Code for TNNLS paper "Beyond Homophily and Homogeneity Assumption: Relation-based Frequency Adaptive Graph Neural Networks"
Python
9
star
9

MD-GNN

Code for NCAA paper "Multi-level Disentanglement Graph Neural Network"
Python
8
star
10

PSC-CPI

Code for AAAI 2024 paper "PSC-CPI: Multi-Scale Protein Sequence-Structure Contrasting for Efficient and Generalizable Compound-Protein Interaction Prediction"
Python
8
star
11

DCV

Code for TNNLS paper "Deep Clustering and Visualization for End-to-End High Dimensional Data analysis"
Python
7
star
12

Homophily-Enhanced-Self-supervision

Code for TNNLS paper "Homophily-Enhanced Self-supervision for Graph Structure Learning: Insights and Directions"
Python
7
star
13

L2A

Code for ECML-PKDD 2023 paper "Learning to Augment Graph Structure for both Homophily and Heterophily Graphs"
Python
4
star
14

LirongWu

1
star
15

lirongwu.github.io

HTML
1
star
16

TGS

Code for TKDE paper "A Teacher-Free Graph Knowledge Distillation Framework with Dual Self-Distillation"
Python
1
star