This is the code associated with the submission "Mining Long Short-Term Evolutionary Patterns for Temporal Knowledge Graph Reasoning" under review at ICPR 2024.
All the processed datasets we used in the paper can be downloaded at Baidu Yun(password:6cha). Put datasets in the folder 'data' to run experimments.
To run LSEN on YAGO for link prediction task:
python main.py -d YAGO --history-len 1 --lambdax 1.0 --graph-layer 6
To run LSEN on WIKI for link prediction task:
python main.py -d WIKI --history-len 1 --lambdax 1.0 --graph-layer 6
To run LSEN on ICEWS18 for link prediction task:
python main.py -d ICEWS18 --history-len 3 --lambdax 2.0 --graph-layer 2
To run LSEN on ICEWS14 for link prediction task:
python main.py -d ICEWS14 --history-len 3 --lambdax 2.0 --graph-layer 2 --use-valid false --max-epochs 15
To run LSEN on GDELT for link prediction task:
python main.py -d GDELT --history-len 3 --lambdax 2.0 --graph-layer 2--dataset the dataset to use (YAGO, WIKI, ICEWS18, ICEWS14, or GDELT)
--device the device to use
--batch-size batch size
--max-epochs maximum epochs
--valid-epochs validation epochs
--alpha alpha for nceloss
--lambdax the hyperparameter lambda
--history-len the time window size
--mode offline or online setting
--graph-layer number of GNN layers
--embedding-dim embedding dimension of entities and relations
--lr learning rate
--weight-decay weight decay ratio
--dropout dropout rate
--grad-norm norm to clip gradient to
--filtering filtering setup
--only-eva whether only evaluation on test set
--use-valid whether using validation set
--model-dir model directory
--save-dir save directory
--eva-dir saved dir of the testing model