BoxEL: Box EL++ Embedding

This is the official pytorch implementation of the BoxEL for the description logic EL++. BoxEL++ is a geometric approach based on box embedding to generate embeddings for the description logic EL++

Requiriments

You need CUDA installed to use a GPU, and need to install python libraries with:

pip install -r requirements.txt

Data

We have preprocessed all the data in /data directory. In particular, we have normalized the ontologies into normal forms and splited the data into train/valid/test sets.

For original data, refer https://bio2vec.cbrc.kaust.edu.sa/data/elembeddings/el-embeddings-data.zip for protein-protein interaction and https://github.com/kracr/EmELpp for subsumption reasoning.

How to run Box EL++

We provide two ways to run our BoxEL++ embeddings: python scripts and jupyter notebooks. Since different dataset/tasks have different features, we provide specific files for each datasts/tasks

Jupyter notebooks

We provided specific jupyter notebooks files for all the tasks used in our paper.

e.g, to run and visualize our family domain example, simply open and run

./notebooks/BoxEL-ToyFamily.ipynb

You could get the the following results

Python scripts

We also provided python scripts to run the tasks.

e.g. to run BoxEL on Gene Ontology, simply run

python scripts/BoxEL-GO.py 

The settings of the used hyperparameters is given in the python files.

Pretrained models

We provided some pretrained models in ./models/