An implementation for the paper--Efficient Learning for Billion-scale Heterogeneous Information Networks.
The three datasets used in the paper (PubMed, Yelp and DBLP) can be downloaded from here. In addition, the OGB-MAG240M dataset can be found here. Please place the downloaded datasets in the ../data.
To conduct the experiments, please execute main.py in each folder (Node Classification, Link Prediction, and MAG240M). Hyperparameters can be explored within the main.py, and here are the ones we used.
| Task | Dataset | K | learning rate | dropout | hidden dimension | layers | batch size | |
|---|---|---|---|---|---|---|---|---|
| Node Classification | PubMed | 0.7 | 20 | 1e-3 | 0.4 | 256 | 4 | 3000 |
| Yelp | 0.7 | 20 | 3e-4 | 0.5 | 256 | 4 | 3000 | |
| DBLP | 0.7 | 20 | 5e-4 | 0.5 | 512 | 5 | 3000 | |
| OGB-MAG240M | 0.7 | 25 | 3e-4 | 0.4 | 512 | 5 | 5000 | |
| Link Prediction | PubMed | 0.1 | 20 | 3e-4 | 0.5 | 256 | 4 | 40 |
| Yelp | 0.1 | 20 | 3e-4 | 0.5 | 256 | 4 | 100 | |
| DBLP | 0.7 | 20 | 5e-4 | 0.5 | 512 | 5 | 1000 |