WSDM 2022 TGP - AntGraph

Source code for "An Effective Graph Learning based Approach for Temporal Link Prediction: The First Place of WSDM Cup 2022"

WSDM Cup 2022

WSDM Cup Website link

Link to this challenge

Datasets

Datasets A

• train_csvs/edges_train_A.csv
• train_csvs/node_features.csv
• train_csvs/edge_type_features.csv
• test_csvs/input_A_initial.csv
• test_csvs/input_A.csv
• test_csvs/input_A_final.csv

Datasets B

• train_csvs/edges_train_B.csv
• test_csvs/input_B_initial.csv
• test_csvs/input_B.csv
• test_csvs/input_B_final.csv

Usage

1. Get LINE embedding

We use LINE to generate a representation for the nodes of the constructed graph, and the heterogeneous edge information is temporarily discarded.

For dataset B, src_id and dst_id refer to different things, that is, they belong to heterogeneous nodes. Therefore, when constructing the LINE representation, we made a transformation for dst_id unification to distinguish between User and Item. It's dst_id = dst_id + dst_offset.

dataset_A_graph_src_dst_id.emb and dataset_B_graph_src_dst_id.emb in the code are line embedding files.

Quick guide

Download the *.emb from this link (OneDrive) and place it in the ./embs folder.

1.1. build input data for LINE

cd src
python ./generate_for_line_emb.py --dataset A
python ./generate_for_line_emb.py --dataset B

1.2. generate LINE embedding

The -threads 64 can be changed according to the actual situation.

We use the linux version of the program to generate the embeddings.

Dataset A

cd src/LINE-master/(linux or windows)

./line -train ../../../embs/dataset_A_graph_src_dst_id.txt -output ../../../embs/dataset_A_graph_src_dst_id.emb -size 200 -order 1 -negative 1 -samples 1000 -threads 64

Settings

Order: 1
Samples: 1000M
Negative: 1
Dimension: 200
Initial rho: 0.025000
----------------------
Number of edges: 54090536
Number of vertices: 19442

Dataset B

cd src/LINE-master/(linux or windows)

./line -train ../../../embs/dataset_B_graph_src_dst_id.txt -output ../../../embs/dataset_B_graph_src_dst_id.emb -size 200 -order 1 -negative 5 -samples 1000 -threads 64

Settings

Order: 1
Samples: 1000M
Negative: 5
Dimension: 200
Initial rho: 0.025000
----------------------
Number of edges: 8278431
Number of vertices: 1304045

2. All Pipeline

AntGraph_solution.ipynb

Dependencies

numpy==1.21.0
seaborn==0.11.2
lightgbm==3.1.1
xgboost==1.3.3
catboost==0.26.1
pandas==1.1.5
tqdm==4.62.2
scikit-learn==0.24.2
matplotlib==3.3.4
wandb==0.12.3

References

Please cite this repository using the following reference:

@article{zhao2022effective,
  title={An Effective Graph Learning based Approach for Temporal Link Prediction: The First Place of WSDM Cup 2022},
  author={Zhao, Qian and Yang, Shuo and Hu, Binbin and Zhang, Zhiqiang and Wang, Yakun and Chen, Yusong and Zhou, Jun and Shi, Chuan},
  journal={arXiv preprint arXiv:2203.01820},
  year={2022}
}