HINormer: Representation Learning On Heterogeneous Information Networks with Graph Transformer

We provide the implementaion of HINormer based on the official PyTorch implementation of HGB(https://github.com/THUDM/HGB)

1. Descriptions

The repository is organised as follows:

• dataset/: the original data of four benchmark dataset.
• run.py: multi-class node classificaiton of HINormer.
• run_multi.py: multi-label node classification of HINormer on IMDB.
• model.py: implementation of HINormer.
• utils/: contains tool functions.
• HGB-output/: contains test files on HGB.

2. Requirements

• Python==3.9.0
• Pytorch==1.12.0
• Networkx==2.8.4
• numpy==1.22.3
• dgl==0.9.0
• scikit-learn==1.1.1
• scipy==1.7.3

3. Running experiments

We train our model using NVIDIA TITAN Xp GPU with CUDA 10.2.

For node classification with offline evaluation:

• python run.py --dataset DBLP --len-seq 50 --dropout 0.5 --beta 0.1 --temperature 2
• python run_multi.py --dataset IMDB --len-seq 20 --beta 0.1 --temperature 0.1
• python run.py --dataset Freebase --num-gnns 3 --len-seq 30 --num-layers 3 --dropout 0 --beta 0.5 --temperature 0.2
• python run.py --dataset AMiner --len-seq 80 --num-gnns 3 --num-layers 4 --temperature 0.5

For node classification with online evaluation on HGB:

• python run.py --dataset DBLP-HGB --len-seq 50 --num-heads 2 --dropout 0.5 --beta 0.1 --temperature 0.1 --mode 1
• python run_multi.py --dataset IMDB-HGB --len-seq 150 --beta 0.5 --temperature 1 --mode 1

And we provide our test files on DBLP-HGB and IMDB-HGB in 'HGB-output/'.

For reproducing our results in the paper and applying HINormer to other datasets, you need to tune the values of key parameters like 'num-gnns','num-layers','len-seq', 'dropout', 'temperature' and 'beta' in your experimental environment.

4. Citation