T-GAP

This is an official implementation of T-GAP: Learning to Walk across Time for Temporal Knowledge Graph Completion (KDD2021).

Requirements

• python >= 3.7
• torch >= 1.4.0
• numpy == 1.17.2
• dgl-cu101 == 0.4.2
• tqdm == 4.36.1
• ipdb == 0.12.3

You can install the requirements using pip install -r requirements.txt.

Datasets, Pre-processing, and Pretrained Model

We provide a shell script to download the datasets and pre-process them. The datasets used in the paper are ICEWS14, ICEWS05-15, and Wikidata11k. Go to the data folder, and type sh preprocess.sh. The script downloads, and pre-processes the official distribution of the 3 public datasets provided by Garcia-Duran et al.

We also provide the version of T-GAP trained on ICEWS14 dataset. The checkpoint file can be found at results/checkpoint/trained_icews14.tar.

Simple Demo

You can run a demo script to evaluate the trained model on ICEWS14, by running sh scripts/demo.sh in the project root.
The experimental results of the trained model are as follows:

MRR	Hits@1	Hits@3	Hits@10
0.610	50.9	67.7	79.0

Note that we have used 1 Tesla V100 GPU for all experiments.

Training

We provide additional scripts for each dataset that trains T-GAP with the same hyperparameter configurations as used in the paper. The scripts can be run as follows:

• ICEWS14: sh scripts/icews14.sh
• ICEWS05-15: sh scripts/icews05-15.sh
• Wikidata11k: sh sh scripts/wikidata11k.sh

Running these files will automatically create checkpoint files and tensorboard log in results/checkpoint/${DATASET}, and results/log/${DATASET} respectively. For a tutorial on how to use tensorboard, refer to Pytorch Tensorboard Tutorial.

To train with custom hyperparameters, run main.py in the project root folder as follows:

python main.py \
  --run=${DIR_NAME} \
  --dataset=${DATASET_NAME} \
  --device=${cuda/cpu} \
  --seed=... \
  --epoch=... \
  --lr=... \
  --grad_clip=... \
  --node_dim=... \
  --num_in_heads=... \
  --num_out_heads=... \
  --num_step=... \
  --num_sample_from=... \
  --max_num_neighbor=... \

Option	Description	Default
run	Output dir name (automatically created under results/checkpoint, results/log)	Current Time
dataset	Dataset to train (among data/icews14_aug, data/icews05-15_aug, data/wikidata11k_aug)	data/icews14_aug
device	Torch device to use (cuda or cpu)	cuda
seed	Random seed	999
epoch	Number of training epochs	10
lr	Learning Rate	5e-4
grad_clip	Gradient clipping norm	3
node_dim	Embedding dimension	100
num_in_heads	Number of PGNN, SGNN attention heads	5
num_out_heads	Number of Attention Flow attention heads	5
num_step	Number of propagation steps	3
num_sample_from	Number of core nodes	10
max_num_neighbor	Number of sampled / added edges	100

Evaluation

You can evaluate a trained model, using Mean Reciprocal Rank (MRR), Hits@1/3/10 on the test set.
For evaluation, you need to locate the checkpoint file generated during training. To the same command used in the training, add --test and --ckpt=${CKPT_DIR} to evaluate.

python main.py
  --test \
  --ckpt=${CKPT_DIR} \
  --other_options (Same configuration with training)

References

Garcia-Duran. A, Dumancic. S, and Niepert. M, Learning Sequence Encoders for Temporal Knowledge Graph Completion. EMNLP 2018