This work is composed of two component, namely encoder and decoder. The encoder is mainly a procedure of feature gather by Hamming distance. We first introduce it to KGC task. The decoder is the famous facotrization model RESCAL, this is why our model named HRSCAL. However, our model not report the metrics of Hit@N, MR and MRR. We report the metric of AUC. The results are as follows.
| Model | Kinships | Nations | UMLS |
|---|---|---|---|
| CP | 0.9400 | 0.8300 | 0.9500 |
| IRM | 0.6600 | 0.7500 | 0.7000 |
| BCTF | 0.9000 | N/A | 0.9800 |
| RESCAL | 0.9500 | 0.8400 | 0.9800 |
| Linear+Reg | 0.9399 | - | 0.8822 |
| Quad+Reg | 0.9389 | - | 0.8811 |
| Linear+Constraint | 0.9287 | - | 0.8018 |
| Quad+Constraint | 0.9384 | - | 0.9107 |
| HRESCAL2(ours) | 0.9991 | 0.9782 | 0.9923 |
| HRESCAL3(ours) | 0.9986 | 0.9604 | 0.9997 |
| HRESCAL3(ours) | 0.9986 | 0.9604 | 0.9997 |
| Improved2 | 5.25% | 16.45% | 1.26% |
| Improved3 | 5.12% | 14.33% | 2.01% |
| Model | Countries S1 | Countries S2 | Countries S3 |
|---|---|---|---|
| HolE | 0.9970 | 0.7720 | 0.6970 |
| ComplEx | 0.9977 | 0.9075 | 0.5474 |
| NTPλ | 1.0000 | 0.9304 | 0.7726 |
| MINERVA | 1.0000 | 0.9304 | 0.7726 |
| NeuralLP | 1.0000 | 0.7510 | 0.9220 |
| GNTP | 1.0000 | 0.9348 | 0.9127 |
| HRESCAL2(ours) | 0.9990 | 0.9987 | 0.9982 |
| HRESCAL3(ours) | 0.9963 | 0.9940 | 0.9909 |
| Improved2 | -0.10% | 7.34% | 8.57% |
| Improved3 | -0.37% | 6.84% | 8.57% |
| Model | FB15K | FB15K237 |
|---|---|---|
| RuleN | - | 0.9225 |
| GNN | - | 0.9337 |
| RESCAL | - | 0.9761 |
| Non Neg RESCAL | - | 0.9781 |
| TransE | - | 0.5084 |
| DisMult | - | 0.7028 |
| ComplEx | - | 0.6764 |
| Linear+Reg | - | 0.9649 |
| Quad+Reg | - | 0.9720 |
| Linear+Constraint | - | 0.8000 |
| Quad+Constraint | - | 0.9459 |
| HRESCAL2(ours) | 0.9907 | 0.9722 |
| Improved2 | - | -0.61% |
| Dataset | Model | Rank 10 | Rank 20 | Rank 40 |
|---|---|---|---|---|
| Kinships | RESCAL | 10.76 | 11.06 | 12.08 |
| HRESCAL2 | 2.38 | 2.45 | 2.51 | |
| Nations | RESCAL | 27.39 | 27.33 | 27.83 |
| HRESCAL2 | 3.46 | 3.63 | 3.77 | |
| UMLS | RESCAL | 28.40 | 29.15 | 29.96 |
| HRESCAL2 | 4.84 | 5.08 | 5.24 |
Our decoder is RESCAL, this part of code is from the RESCAL, the details of this model, please refer to the link.
When we use the model HRESCAL, we should first convert the data into .mat, then change the path to your file on your equipment, then your can run it.
You can use the following command:
python yourpath/getMat_txt.pyor
python yourpath/getMat_tsv.pyWe just provide the txt and tsv convertor. Other convertors are similar.
python yourpath/KB_kinships_best.pyWe provide all of our running code, you can run them at your will.
rescal.py is licensed under the GPLv3 http://www.gnu.org/licenses/gpl-3.0.txt