DegUIL [ECML-PKDD, 2023]

The repository provides PyTorch code and datasets for our paper "DegUIL: Degree-aware Graph Neural Networks for Long-tailed User Identity Linkage", which is published in 23'ECML-PKDD.

1. Requirements

• PyTorch >= 1.8.1+cu111
• Python 3.8.5
• Numpy 1.19.5
• nltk 3.6.1
• tqdm 4.60.0

To install, run pip install -r requirements.txt. Our code is run in GPU by default (CUDA 11.1+ here) , you can change the device into CPU if only CPU is available.

2. Repository Strucutre

• datasets/: contains the datasets FT(Foursquare-Twitter) and DBLP(DBLP17-DBLP19), which come from open source. The bestEmbs folder stores the mapped embeddings of two networks when getting the best performance.
• layers/: model layers
• models/: main modules of our model DegUIL
• utils:/ tool functions for processing data and logging
• config.py: hyperparameters
• DegUIL.py: main file

Some codes refer to https://github.com/shuaiOKshuai/Tail-GNN

3. Usage

# FT dataset
python Node2Vec/run_node2vec.py -- dataset FT   # get emb_n2v1.pkl
python DegUIL.py --dataset FT --mu 0.001 --lr 5e-4

# DBLP dataset
python Node2Vec/run_node2vec.py -- dataset DBLP
python DegUIL.py --dataset DBLP --mu 0.01 --lr 1e-3

or

bash run.sh

Note:

Each dataset in datasets/ includes two original adjacency matrices adj_s.pkl and adj_t.pkl, their corresponding node embeddings generated by node2vec, are saved as emb_n2v1.pkl. To run the code on your own datasets, please refer to utils/data_process.py to process the corresponding datasets into the input format.

4. Citation

@inproceedings{long2023deg,
  title={DegUIL: Degree-aware Graph Neural Networks for Long-tailed User Identity Linkage},
  author={Meixiu Long, Siyuan Chen, Xin Du, and Jiahai Wang},
  booktitle = {{ECML/PKDD}},
  volume = {14174},
  pages = {122-138},
  publisher = {Springer},
  year = {2023}
}