The code is about 《NOTE: Solution for KDD-Cup 2021 WikiKG90M-LSC》.
ogb>=1.3.0
torch>=1.7.0
paddle>=2.0.0
pgl>=2.1.2
python ./feature/walk_probability/h2r.py
python ./feature/walk_probability/h2t.py
python ./feature/walk_probability/r2t.py
python ./feature/walk_probability/r2h.py
python ./feature/walk_probability/t2r.py
python ./feature/walk_probability/t2h.py
python ./feature/dump_feat/1_rrt_feat.py
python ./feature/dump_feat/2_h2t_t2h_feat.py
python ./feature/dump_feat/3_t2h_h2t_feat.py
python ./feature/dump_feat/4_h2t_h2t_feat.py
python ./feature/dump_feat/5_hht_feat_byprob.py
python ./feature/dump_feat/6_r2t_h2r_feat.py
python ./feature/dump_feat/7_r2t_feat.py
python ./feature/dump_feat/8_rrh_feat.py
python ./feature/dump_feat/9_rt_feat.py
python ./feature/dump_feat/10_ht_feat.py
This will give you the manual features in feature_output
.
Please refer to Graph4KG for paddle version code.
The DGL-KE version code is in model_dglke.
sh ./model_dglke/run_train_all.sh
sh ./model_dglke/run_infer_all.sh
This will give you the manual features in model_output
.
For the trained TransE and RotatE models, we use post smooothing technology to further improve the effect
sh post_smoothing/run_transe.sh
sh post_smoothing/run_rotate.sh
# This will automatically ensemble results from ./model_output/, ./feature_output/ and generate t_pred_wikikg90m.npz
python ensemble.py
This will give you NOTE value in the performance table below
Model | Valid ACC |
---|---|
Final Ensemble(w/ manual feature) | 0.9797 |