/DSKG

Primary LanguageJupyter Notebook

DSKG

A Deep Sequential Model for Knowledge Graph Completion

Requirements

  • python 3.x
  • tensorflow 1.x
  • numpy, pandas
  • jupyter

Running

  1. unpack the data.tar.gz, which includes all of three datasets.

  2. run jupyter: jupyter notebook

  3. open runDSKG.ipynb & run all cells (Kernel -> Restart & Run All)

You can also directly click runDSKG.ipynb in this page to preview the results we have run.

Entity Prediction Results

FB15K-237

Models Hits@1 Hits@10 MRR MR
TransE (our) 13.3 40.9 22.3 315
TransR (our) 10.9 38.2 19.9 417
PTransE (our) 21.0 50.1 31.4 299
DISTMULT 15.5 41.9 24.1 254
NLFeat - 41.4 27.2 -
ComplEx 15.2 41.9 24.0 248
NeuralLP - 36.2 24.0 -
ConvE 23.9 49.1 31.6 246
InverseModel 0.4 1.2 0.7 7,124
DSKG (cascade) 20.5 50.1 30.3 842
DSKG 24.9 52.1 33.9 175

FB15K

Models Hits@1 Hits@10 MRR MR
TransE (our) 30.5 73.7 45.8 71
TransR (our) 37.7 76.7 51.9 84
PTransE (our) 63.8 87.2 73.1 59
DISTMULT 54.6 82.4 65.4 97
NLFeat - 87.0 82.1 -
ComplEx 59.9 84.0 69.2 -
NeuralLP - 83.7 76.0 -
ConvE 67.0 87.3 74.5 64
InverseModel 74.3 78.6 75.9 1,563
DSKG (cascade) 64.9 87.7 73.0 151
DSKG 75.3 90.2 80.9 30

WN18

Models Hits@1 Hits@10 MRR MR
TransE (our) 27.4 94.4 57.8 431
TransR (our) 54.8 94.7 72.6 415
PTransE (our) 87.3 94.2 90.5 516
DISTMULT 72.8 93.6 82.2 902
NLFeat - 94.3 94.0 -
ComplEx 93.6 94.7 94.1 -
NeuralLP - 94.5 94.0 -
ConvE 93.5 95.5 94.2 504
InverseModel 75.7 96.9 85.7 602
DSKG (cascade) 93.9 95.0 94.3 959
DSKG 94.2 95.2 94.6 337