We combine node labels with scores calculated by local similarity measures and use a simple MLP for the link prediction task which can obtain good performance.
Python>=3.6
Pytorch>=1.4
torch-geometric>=1.6.0
ogb>=1.3.1
python generate_feature.pypython train.py --sim all
We conduct the experiments for 10 times with the random seed 0~9 and results are listed below:
| Model | Test Hits@100 | Val Hits@100 |
|---|---|---|
| MLP+CN | 0.3064±0.0116 | 0.3161±0.0070 |
| MLP+RA | 0.4896±0.0048 | 0.4794±0.0029 |
| MLP+AA | 0.3459±0.0033 | 0.3454±0.0029 |
| MLP+RA&CN&AA | 0.5062±0.0035 | 0.4906±0.0029 |
We use multiple anchor sets selected from random sampling to encode distance information for edges on graph. We also modify the aggregation stage of GraphSAGE to incorporate edge information.
To get the best performance, run:
python link_pred_ddi_graphsage_edge.py --node_emb 512 --hidden_channels 512 --num_samples 3| Model | Test Hits@20 | Val Hits@20 |
|---|---|---|
| GraphSAGE+Edge Attr(k=1) | 0.8633±0.0313 | 0.7916±0.0324 |
| GraphSAGE+Edge Attr(k=3) | 0.8781±0.0474 | 0.8044±0.0404 |
| GraphSAGE+Edge Attr(k=5) | 0.8527±0.0247 | 0.7839±0.0278 |