Awesome
RSN
Lingbing Guo, Zequn Sun, Wei Hu*. Learning to Exploit Long-term Relational Dependencies in Knowledge Graphs. In: ICML 2019
INSTALLATION
-
Please first install Python 3.5+, and then unpack data.7z.
-
Type <code>pip install -r requirements</code> in shell to install required packages. Note that, when using Tensorflow 1.2+, the learning rate has to be re-adjusted. We suggest using tensorflow-gpu = 1.1.
RUNNING
-
Run jupyter by typing <code>jupyter notebook</code> in shell.
-
In the opened browser, click RSN4EA.ipynb for entity alignment, or RSN4KGC.ipynb for KG completion.
-
The files RSN4EA.ipynb and RSN4KGC.ipynb record the latest results on DBP-WD (normal) and FB15K, respectively.
-
You can also click 'Toolbar -> Kernel -> Restart&Run All' to run these two experiments.
DATA
-
Limited by the space, we only uploaded FB15K for KG completion. For WN18 and FB15K-237, you can easily download them from the Internet.
-
Change options.data_path or other options.* to run RSN on different datasets with different settings.
-
For RSN4KGC.ipynb, we adopt a matrix filter method for evaluation, which may use more than 64GB memory.
-
For entity alignment, V1 denotes the normal datasets, and V2 denotes the dense ones. Please use first 10% data of <code> ref_ent_ids</code> for validation.
-
entity-alignment-full-data.7z provides a complete version of the EA datasets, including attribute files and datasets with different proportions.
CITATION
If you find our work useful, please kindly cite it as follows:
@inproceedings{RSN,
author = {Lingbing Guo and Zequn Sun and Wei Hu},
title = {Learning to Exploit Long-term Relational Dependencies in Knowledge Graphs},
booktitle = {ICML},
pages = {2505--2514},
year = {2019}
}