OpenHGNN/openhgnn/output/GTN

• GTN[NeurIPS2019] & fastGTN
• • • How to run
    • • • Candidate dataset:
    • Performance
    • TrainerFlow: node_classification
    • • model
    • Dataset
    • • acm4GTN/imdb4GTN
    • Hyper-parameter specific to the model
    • Related API in DGL
  • More
  • • • Contirbutor
      • If you have any questions,

GTN[NeurIPS2019] & fastGTN

Paper: [GTN] Graph Transformer Networks

Extension Paper: [fastGTN] Graph Transformer Networks: Learning Meta-path Graphs to Improve GNNs

Code from author: https://github.com/seongjunyun/Graph_Transformer_Networks

How to run

Clone the OpenHGNN

# Run GTN
python main.py -m GTN -t node_classification -d acm4GTN -g 0 --use_best_config
# Run the fastGTN
python main.py -m fastGTN -t node_classification -d acm4GTN -g 0 --use_best_config

If you do not have gpu, set -gpu -1.

Candidate dataset:

​ acm4GTN/imdb4GTN/dblp4GTN

Performance

Node classification

Node classification(F1 score)	acm4GTN	imdb4GTN	dblp4GTN
paper[GTN]	92.68	60.92	94.18
OpenHGNN[GTN]	Macro: 92.03 Micro: 92.00	Macro: 56.97 Micro: 58.61	87.33(OOM on Tesla T4(16GB), cpu result)
OpenHGNN[fastGTN]	Macro: 92.92 Micro: 92.85	Macro: 60.62 Micro: 62.59	Macro: 90.39 Micro: 91.39

TrainerFlow: node_classification

The model is trained in semi-supervisied node classification.

model

• transform_relation_graph_list
  • Extract a graph list where every graph just contains a relation.
• GTLayer
  • Contain GTConv
  • Contain the product of the adjacency matrices of two graphs getting from GTConv.
• GTConv
  • Create a weighted graph whose adjacency matrix is the sum of the adjacency matrices of the given graph list.

Dataset

Supported dataset: acm4GTN, imdb4GTN, dblp4ACM

Note: Every node in dataset should have the same features dimension.

Dataset Description

acm4GTN/imdb4GTN

We process the acm dataset given by HAN. It saved as dgl.heterograph and can be loaded by dgl.load_graphs

You can download the dataset by

wget https://s3.cn-north-1.amazonaws.com.cn/dgl-data/dataset/acm4GTN.zip
wget https://s3.cn-north-1.amazonaws.com.cn/dgl-data/dataset/imdb4GTN.zip

Or run the code mentioned above and it will download automaticlly.

Hyper-parameter specific to the model

num_channels = 2 # number of channel
num_layers = 3 # number of layer
adaptive_lr_flag = True # use different learning rate for weight in GTLayer.

Best config can be found in best_config

Related API in DGL

dgl.adj_product_graph which is equivalent SpSpMM.

dgl.adj_sum_graph

GraphConv, EdgeWeightNorm

More

Contirbutor

Tianyu Zhao[GAMMA LAB]

If you have any questions,

Submit an issue or email to tyzhao@bupt.edu.cn.