Representation Learning of Knowledge Graphs with Embedding Subspaces

Table 3

Prediction results with different training set sizes on FB15K-237.

		Mean rank		HITs@10 (%)
Training set size	Method	Raw	Filtered	Raw	Filtered
Train = 5000	TransE	4559	4544	13.6	14.0
	TransH	4568	4554	13.4	13.9
	ProjE	1155	1141	29.6	41.6
	Sub-ProjE	447	434	19.5	20.3
Train = 10000	TransE	2902	2886	18.9	19.9
	TransH	2972	2955	18.5	19.5
	ProjE	614	599	38.2	52.2
	Sub-ProjE	377	362	23.7	24.9
Train = all	TransE	297	219	42.9	58.3
	TransH	293	217	42.5	58.0
	ProjE	164	98	40.8	64.9
	Sub-ProjE	313	246	25.7	35.3