Heterogeneous Network Representation Learning: Benchmark with Data and Code

Citation

Please cite the following work if you find the data/code useful.

@article{yang2020heterogeneous,
  title={Heterogeneous Network Representation Learning: A Unified Framework with Survey and Benchmark},
  author={Yang, Carl and Xiao, Yuxin and Zhang, Yu and Sun, Yizhou and Han, Jiawei},
  journal={TKDE},
  year={2020}
}

Contact

Please contact us if you have problems with the data/code, and also if you think your work is relevant but missing from the survey.

Yuxin Xiao (yuxinx2@illinois.edu), Carl Yang (yangji9181@gmail.com)

Guideline

Stage 1: Data

We provide 4 HIN benchmark datasets: DBLP, Yelp, Freebase, and PubMed.

Each dataset contains:

• 3 data files (node.dat, link.dat, label.dat);
• 2 evaluation files (link.dat.test, label.dat.test);
• 2 description files (meta.dat, info.dat);
• 1 recording file (record.dat).

Please refer to the Data folder for more details.

Stage 2: Transform

This stage transforms a dataset from its original format to the training input format.

Users need to specify the targeting dataset, the targeting model, and the training settings.

Please refer to the Transform folder for more details.

Stage 3: Model

We provide 13 HIN baseline implementaions:

• 5 Proximity-Preserving Methods (metapath2vec-ESim, PTE, HIN2Vec, AspEm, HEER);
• 4 Message-Passing Methods (R-GCN, HAN, MAGNN, HGT);
• 4 Relation-Learning Methods (TransE, DistMult, ComplEx, ConvE).

Please refer to the Model folder for more details.

Stage 4: Evaluate

This stage evaluates the output embeddings based on specific tasks.

Users need to specify the targeting dataset, the targeting model, and the evaluation tasks.

Please refer to the Evaluate folder for more details.