Knowledge Graph Reasoning

This page contains benchmarks of knowledge graph reasoning methods.

We use the filtered ranking protocol for knowledge graph reasoning. For a test triplet \(\langle h, r, t\rangle\), we rank it against all negative triplets \(\langle h, r, t'\rangle\) or \(\langle h', r, t\rangle\) that do not appear in the knowledge graph. We report the mean rank (MR), mean reciprocal rank (MRR) and HITS at K (HITS@K) over the test set.

FB15k

FB15k	MR	MRR	HITS@1	HITS@3	HITS@10
TransE	47	0.682	0.585	0.752	0.839
DistMult	43	0.674	0.578	0.743	0.836
ComplEx	38	0.673	0.577	0.738	0.842
SimplE	45	0.689	0.603	0.747	0.836
RotatE	58	0.764	0.703	0.806	0.867
KBGAT	65	0.601	0.489	0.680	0.790
NeuralLP	928	0.660	0.607	0.692	0.757

FB15k-237

FB15k-237	MR	MRR	HITS@1	HITS@3	HITS@10
TransE	189	0.307	0.211	0.343	0.500
DistMult	212	0.271	0.187	0.297	0.436
ComplEx	206	0.271	0.187	0.297	0.437
SimplE	212	0.268	0.185	0.293	0.433
RotatE	202	0.316	0.223	0.349	0.507
KBGAT	222	0.267	0.175	0.298	0.448
NeuralLP	1943	0.259	0.196	0.281	0.384

WN18

WN18	MR	MRR	HITS@1	HITS@3	HITS@10
TransE	297	0.738	0.637	0.809	0.927
DistMult	668	0.744	0.592	0.888	0.932
ComplEx	692	0.744	0.593	0.888	0.933
SimplE	675	0.929	0.922	0.934	0.940
RotatE	406	0.947	0.941	0.951	0.955
KBGAT	231	0.743	0.636	0.830	0.933
NeuralLP	7893	0.740	0.723	0.750	0.771

WN18RR

WN18RR	MR	MRR	HITS@1	HITS@3	HITS@10
TransE	3071	0.220	0.015	0.389	0.529
DistMult	4389	0.447	0.405	0.460	0.541
ComplEx	4635	0.449	0.408	0.458	0.538
SimplE	4935	0.472	0.436	0.488	0.540
RotatE	3917	0.476	0.432	0.494	0.563
KBGAT	3737	0.206	0.015	0.366	0.496
NeuralLP	20950	0.346	0.320	0.354	0.400

Hetionet

Hetionet	MR	MRR	HITS@1	HITS@3	HITS@10
TransE	1088	0.162	0.102	0.173	0.284
DistMult	941	0.187	0.128	0.199	0.304
ComplEx	800	0.235	0.166	0.255	0.374
SimplE	893	0.194	0.134	0.207	0.313
RotatE	744	0.257	0.185	0.282	0.403
KBGAT	1713	0.058	0.023	0.055	0.130
NeuralLP	4017	0.175	0.128	0.182	0.273