README

We use Anonymous Github for the double-blind reviewing policy.

Environment

• python 3.6.13

• cuda10.0.130

• pytorch 1.8.0

Datasets

• ICEWS14 and ICEWS05-15: Learning Sequence Encoders for Temporal Knowledge Graph Completion

• Wikidata12k: HyTE: Hyperplane-based Temporally aware Knowledge Graph Embedding

Test

For ICEWS14 and ICEWS05-15, the coarse time granularity $\Delta t$ is selected from {year, quarter, month}, the learning rate $lr$ from {0.0005, 0.001, 0.005, 0.001}, the dimension of entity embedding $d_e$ from {30, 50, 100, 150}, the dimension of relation embedding $d_r$ from {20, 30, 50, 100}, and the time smoothing weight $\alpha$ from {1, 0.1, 0.01, 0.001}.

For Wikidata12k, the coarse time granularity $\Delta t$ is selected from {century, decade, 5 years}, and the set of candidate values for $lr$, $d_e$, $d_r$, and $\alpha$ are the same as that in ICEWS.

The final parameters selected:

	coarse_grain	batch_size	label_smoothing	lr	dr	entity_dim	rel_dim	alpha	dropout1	dropout2
ICEWS14	month	128	0.1	0.001	1	200	50	0.01	0.5	0.5
ICEWS05-15	quarterly	128	0.1	0.001	1	100	50	0.01	0.5	0.5
Wikidata12k	years5	16	0.1	0.001	1	30	20	0.1	0.5	0.5

Firstly, process the data using:

python process_data.py

Then obtain the test results using:

python main.py --dataset="icews14" --coarse_grain="month" --batch_size=128 --label_smoothing=0.1 --lr=0.001 --dr=1.0 --entity_dim=200 --rel_dim=50 --alpha=0.01 --dropout1=0.5 --dropout2=0.5

python main.py --dataset="icews05-15" --coarse_grain="quarterly" --batch_size=128 --label_smoothing=0.1 --lr=0.001 --dr=1.0 --entity_dim=100 --rel_dim=50 --alpha=0.1 --dropout1=0.5 --dropout2=0.5

python main.py --dataset="wikidata12k" --coarse_grain="years5" --batch_size=16 --label_smoothing=0.1 --lr=0.001 --dr=1.0 --entity_dim=30 --rel_dim=20 --alpha=0.1 --dropout1=0.5 --dropout2=0.5