search

awslabs / dgl-ke

High performance, easy-to-use, and scalable package for learning large-scale knowledge graph embeddings.
https://dglke.dgl.ai/doc/
Apache License 2.0
1.28k stars 196 forks source link

Question about TransE HITS@10 on FB15K dataset #69

Closed artiit closed 4 years ago

artiit commented 4 years ago

Hi, I found that your hit@10 result on transE is 0.8x, but in fact the original paper is 0.4x, I am puzzled about this.

aksnzhy commented 4 years ago

Hi. The original paper only uses 50 for the latent dimension size (we use 400). You can see this paper https://arxiv.org/pdf/1903.12287.pdf (Table-2) to compare our results with pytorch-biggraph (using 400 dim_size) Actually if we set dim_size to 400, the HITS@10 result of the original paper implementation is 0.74. You can also see some other open-source tools' results: https://github.com/thunlp/KB2E In KB2E, using dim_size 100 the HITS@10 is 0.702.