Skip to content

Latest commit

 

History

History
113 lines (83 loc) · 3.82 KB

readme.md

File metadata and controls

113 lines (83 loc) · 3.82 KB

Hetergenour Graph Transformers

Are we really making much progress? Revisiting, benchmarking, and refining heterogeneous graph neural networks

  • authors: Qingsong Lv, Ming Ding, Qiang Liu, Yuxiang Chen, Wenzheng Feng, Siming He, Chang Zhou, Jianguo Jiang, Yuxiao Dong, Jie Tang
  • link: https://arxiv.org/abs/2112.14936
  • file structure:
├── cresci-2015
│   ├── HGT.py  # train HGT on cresci-2015
│   ├── SimpleHGN.py # train SimpleHGN on cresci-2015
│   ├── Dataset.py
│   └── layer.py
├── Twibot-20    
│   ├── HGT.py  # train HGT on Twibot-20
│   ├── SimpleHGN.py # train SimpleHGN on cresci-2015
│   ├── Dataset.py
│   └── layer.py
└── Twibot-22
    ├── HGT_sample.py  # train HGT on Twibot-22
    ├── SimpleHGN_sample.py # train SimpleHGN on Twibot-22
    └── layer.py
  • implement details:

  1. We just consider the graph with user as node and their follow relations (e.g. follower & following) as edge due to the limitation of computation resources.

  2. We fix HGT and SimpleHGN's attention head to 1 for simplicity.

  3. To enable the training on large-scale graph such as Twibot-22, we adopt the NeighborLoader in torch geometric with num_neighbors set to 20 and sample for 2 iterations.

  4. Input features for user node is taken from BotRGCN, more details please refer to BotRGCN imlementation

  • Requirements
CUDA==10.2
pytorch-lightning==1.4.9
torch==1.9.1
torch-geometric==2.0.2
torch-cluster==1.5.9
torch-scatter==2.0.9
torch-sparse==0.6.12
torch-spline-conv==1.2.1
sklearn
argparse

How to reproduce:

Please change the path argument into your own path to preprocess files

  • reproduce HGT on cresci-2015 & Twibot-20
CUDA_VISIBLE_DEVICES=0 python HGT.py --batch_size 128 --epochs 200 --path /path/to/preprocess/file
  • reproduce HGT on Twibot-22
CUDA_VISIBLE_DEVICES=0 python HGT.py --batch_size 128 --epochs 200 --path /path/to/preprocess/file
  • reproduce SimpleHGN on cresci-2015 & Twibot-2015
CUDA_VISIBLE_DEVICES=0 python SimpleHGN.py --batch_size 128 --epochs 200 --path /path/to/preprocess/file
  • reproduce SimpleHGN on Twibot-22 dataset
CUDA_VISIBLE_DEVICES=0 python SimpleHGN.py --batch_size 128 --epochs 200 --path /path/to/preprocess/file

Result:

HGT
dataset acc precison recall f1
Cresci-2015 mean 0.9603 0.9480 0.9911 0.9693
Cresci-2015 std 0.0032 0.0049 0.0012 0.0024
Twibot-20 mean 0.8691 0.8555 0.9100 0.8819
Twibot-20 std 0.0024 0.0031 0.0057 0.0019
Twibot-22 mean 0.7491 0.6822 0.2803 0.3960
Twibot-22 std 0.0013 0.0271 0.0260 0.0211
SimpleHGN
dataset acc precison recall f1
Cresci-2015 mean 0.9671 0.9568 0.9929 0.9745
Cresci-2015 std 0.0054 0.0090 0.0040 0.0040
Twibot-20 mean 0.8674 0.8476 0.9206 0.8825
Twibot-20 std 0.0022 0.0046 0.0051 0.0018
Twibot-22 mean 0.7672 0.7257 0.3290 0.4544
Twibot-22 std 0.0027 0.0279 0.0164 0.0042
baseline acc on Twibot-22 f1 on Twibot-22 type tags
HGT 0.7491 0.3960 F T G Graph Neural Networks
SimpleHGN 0.7672 0.4544 F T G Graph Neural Networks