TransERR: Translation-based Knowledge Graph Embedding via Efficient Relation Rotation

The code of TransERR accepted by the LREC-COLING 2024.

🔬 Dependencies

pip install ogb==1.3.6
PyTorch >= 1.0

📜 Hyperparameters

We list the best hyperparameters setting of TransERR on 10 benchmark datasets. $d$, $\epsilon$, $n$, $\alpha$ and $\gamma$ denote the embedding size, the learning rate, the negative sample, the self-adversarial sampling temperature and fixed margin, respectively.

Benchmark	embedding dimension $d$	learning rate $\epsilon$	negative sample $n$	$\alpha$	$\gamma$
ogbl-wikikg2	200	$1e^{-3}$	128	1.0	15
ogbl-wikikg2	100	$1e^{-3}$	128	1.0	8
ogbl-biokg	2000	$5e^{-5}$	128	1.0	20
ogbl-biokg	1000	$1e^{-4}$	128	1.0	15
YAGO3-10	1000	$5e^{-4}$	256	0.5	30
DB100K	1000	$1e^{-4}$	128	0.5	20
FB15K	1500	$1e^{-4}$	256	1.0	26
WN18	1000	$1e^{-3}$	128	0.5	10
FB15K-237	1000	$1e^{-3}$	128	0.5	20
WN18RR	1000	$1e^{-3}$	128	0.5	12
Sports	200	$1e^{-3}$	128	0.5	5
Location	200	$1e^{-3}$	128	0.5	10

🚀ogbl-wikikg2

Training & Evaluation

# Run with default config 
cd wikikg2
bash examples.sh

We have included a set of training logs, available at wikikg2/log, which have been generated using ten different random seeds, designated by numbers between 0 and 9. These logs are intended to aid in reproducing and validating our experimental results.

🚀ogbl-biokg

Training & Evaluation

# Run with default config
cd biokg
bash examples.sh

🚀wn18rr, fb15k-237, yago3-10, db100k, wn18, fb15k, sports and localtion

Training & Evaluation

# Run with default config
cd code
bash best_config.sh