TPNet for TGB Leaderboard

This is an adapted version of TPNet for TGB Learderboard, where the data loading & metric computing is provided by the py-tgb library. The code of TPNet (i.e., TPNet.py) remains the same as its official implementation. We only added a new hyperparameter --not_embedding in load_configs.py to optionally disable the embedding model for ablation studies. Note that this option was set to False when producing the results submitted to the TGB Leaderboard.

Environments

• python=3.9.18
• pytorch=2.0.1
• numpy=1.21.6
• pandas=1.5.3
• scikit-learn==1.3.0
• py-tgb==2.0.0
• wandb=0.18.3
• tqdm
• tabulate

Executing Scripts for Reproducing the TGB Leaderboard Results

Necessary Directories Creating

mkdir logs saved_results saved_models

Model Training

python train_link_prediction.py --prefix std --dataset_name tgbl-wiki --model_name TPNet --gpu 0 --load_best_configs \
--patience 5 --num_epochs 30 --use_random_projection

python train_link_prediction.py --prefix std --dataset_name tgbl-review --model_name TPNet --gpu 0 --load_best_configs \
--patience 5 --num_epochs 30 --use_random_projection

python train_link_prediction.py --prefix std --dataset_name tgbl-coin --model_name TPNet --gpu 0 --load_best_configs \
--patience 1 --num_epochs 10 --use_random_projection

python train_link_prediction.py --prefix std --dataset_name tgbl-comment --model_name TPNet --gpu 0 --load_best_configs \
--patience 1 --num_epochs 10 --use_random_projection

python train_link_prediction.py --prefix std --dataset_name tgbl-flight --model_name TPNet --gpu 0 --load_best_configs \
--patience 1 --num_epochs 10 --use_random_projection

Reloading Model for Evaluation

python evaluate_link_prediction.py --prefix std --dataset_name tgbl-wiki --model_name TPNet --gpu 0 --load_best_configs --use_random_projection

python evaluate_link_prediction.py --prefix std --dataset_name tgbl-review --model_name TPNet --gpu 0 --load_best_configs --use_random_projection

python evaluate_link_prediction.py --prefix std --dataset_name tgbl-coin --model_name TPNet --gpu 0 --load_best_configs --use_random_projection

python evaluate_link_prediction.py --prefix std --dataset_name tgbl-comment --model_name TPNet --gpu 0 --load_best_configs --use_random_projection

python evaluate_link_prediction.py --prefix std --dataset_name tgbl-flight --model_name TPNet --gpu 0 --load_best_configs --use_random_projection