KGE

A Framework of Knowledge Graph Embedding Models (Including TransE, TransH, ConvKB) by tensorflow.

Reference

(1) TransE: Translating Embeddings for Modeling Multi-relational Data
(2) TransH: Knowledge Graph Embedding by Translating on Hyperplanes
(3) ConvKB: A Novel Embedding Model for Knowledge Base Completion Based on Convolutional Neural Network

Results

(1) Fixed embedding dimension of 100.
(2) Only considering filter setting.
(3) Early stopped by Hist@10 reuslt on dev dataset.

FB15k (14951 E + 1345 R)

	MR	MRR	Hist@1	Hist@3	Hist@10
TransE	85.0	0.575	0.446	0.667	0.790
TransH	82.6	0.580	0.452	0.670	0.793
ConvKB	79.2	0.596	0.470	0.686	0.805

python Run_KGE.py --model [TransX] --dataset FB15k --margin 1.0 --l_r 5e-3 --batch_size 10000 --epoches 5000 --earlystop 2

python Run_KGE.py --model ConvKB --dataset FB15k --n_filter 50 --l_r 1e-3 --batch_size 10000 --epoches 500 --earlystop 5

FB15k-237 (14541 E + 237 R)

	MR	MRR	Hist@1	Hist@3	Hist@10
TransE	299.3	0.312	0.225	0.347	0.482
TransH	300.0	0.318	0.230	0.355	0.488
ConvKB	251.7	0.330	0.242	0.365	0.500

python Run_KGE.py --model [TransX] --dataset FB15k-237 --margin 1.0 --l_r 5e-3 --batch_size 10000 --epoches 5000 --earlystop 2

python Run_KGE.py --model ConvKB --dataset FB15k-237 --n_filter 50 --l_r 1e-3 --batch_size 10000 --epoches 500 --earlystop 5

WN18 (40943 E + 18 R)

	MR	MRR	Hist@1	Hist@3	Hist@10
TransE	290.0	0.462	0.106	0.808	0.937
TransH	302.0	0.465	0.105	0.817	0.936
ConvKB	300.2	0.499	0.226	0.743	0.911

python Run_KGE.py --model [TransX] --dataset WN18 --margin 4.0 --l_r 5e-3 --batch_size 3000 --epoches 5000 --earlystop 2

python Run_KGE.py --model ConvKB --dataset WN18 --n_filter 50 --l_r 1e-3 --batch_size 3000 --epoches 500 --earlystop 5

WN18RR (40943 E + 11 R)

	MR	MRR	Hist@1	Hist@3	Hist@10
TransE	3848.1	0.185	0.012	0.319	0.471
TransH	3795.8	0.188	0.011	0.330	0.474
ConvKB	4317.9	0.153	0.018	0.236	0.415

python Run_KGE.py --model [TransX] --dataset WN18RR --margin 4.0 --l_r 5e-3 --batch_size 3000 --epoches 5000 --earlystop 2

python Run_KGE.py --model ConvKB --dataset WN18RR --n_filter 50 --l_r 1e-3 --batch_size 3000 --epoches 500 --earlystop 5

Kinship (104 E + 25 R)

	MR	MRR	Hist@1	Hist@3	Hist@10
TransE	6.7	0.463	0.258	0.607	0.834
TransH	5.3	0.543	0.334	0.711	0.878
ConvKB	3.4	0.690	0.557	0.786	0.936

python Run_KGE.py --model [TransX] --dataset Kinship --margin 0.1 --l_r 5e-3 --batch_size 500 --epoches 5000 --earlystop 2

python Run_KGE.py --model ConvKB --dataset Kinship --n_filter 50 --l_r 1e-3 --batch_size 500 --epoches 500 --earlystop 5

NELL-995 (75492 E + 200 R)

	MR	MRR	Hist@1	Hist@3	Hist@10
TransE	6217.5	0.351	0.277	0.394	0.477
TransH	6271.8	0.362	0.291	0.404	0.491
ConvKB	7733.1	0.295	0.241	0.324	0.384

python Run_KGE.py --model [TransX] --dataset NELL-995 --margin 5.0 --l_r 5e-3 --batch_size 5000 --epoches 5000 --earlystop 2

python Run_KGE.py --model ConvKB --dataset NELL-995 --n_filter 50 --l_r 1e-3 --batch_size 5000 --epoches 500 --earlystop 5

Kinship (135 E + 46 R)

	MR	MRR	Hist@1	Hist@3	Hist@10
TransE	1.8	0.861	0.766	0.952	0.986
TransH	1.6	0.854	0.750	0.950	0.987
ConvKB	1.5	0.881	0.781	0.979	0.991

python Run_KGE.py --model [TransX] --dataset UMLS --margin 0.1 --l_r 5e-3 --batch_size 500 --epoches 5000 --earlystop 2

python Run_KGE.py --model ConvKB --dataset UMLS --n_filter 50 --l_r 1e-3 --batch_size 500 --epoches 500 --earlystop 5

[TransX] from {TransE, TransH}