This repository is the official implementation of the paper Dual Box Embeddings for the Description Logic EL++.

This implementation requires a working installation of PyTorch 1.12.0 (optionally with GPU support for faster training and inference). You will additionally need the following Python packages: joblib==1.1.0, numpy==1.22.3, tqdm==4.64.0, wandb==0.13.9.

To obtain the data, unzip data.zip:

unzip data.zip

Our benchmark for subsumption prediction is included in the prediction subfolder of the folder of the relevant ontology, e.g., the data for GALEN can be found in data/GALEN/prediction. The data is split into training, validation and testing sets in the relevant subfolders, and we include json files that specify the mapping of classes and relations to integers used in the tensor-based representation.

The data for the PPI prediction task can be found in the PPI subfolder. We again include the training/validation/testing splits and the mapping from classes and relations to integers.

The deductive reasoning benchmark data is contained in the inferences subfolder. It consists of the training data in form of the full OWL ontology, and validation and testing sets as json files.

In order to train Box²EL or one of the baseline methods, edit the file train.py (for subsumption prediction and deductive reasoning) or train_ppi.py (for PPI prediction) with the desired combination of method and dataset. For example, to run Box²EL for subsumption prediction on GALEN, you need to:

1. Open the file train.py
2. In the run function, set the task to 'prediction' (or 'inferences' for deductive reasoning)
3. Set the model and desired hyperparameters
4. Run the file

Training should finish within a couple of minutes on a GPU. The best performing model on the validation set will be saved, evaluated on the testing set, and the results will be printed to the console.

We also provide the script run_many_seeds.py, which executes the configuration in train.py five times and reports the average results.

To evaluate trained models, we provide the files evaluate.py and evaluate_ppi.py.

Our model achieves the following performance (combined across normal forms) on subsumption prediction:

PPI prediction:

Approximating deductive reasoning:

• Out of memory. If you run out of memory, this is most likely due to a too large batch size during rank computation. Try decreasing the batch_size default argument in evaluate.py / evaluate_ppi.py.

If you encounter any other issues or have general questions about the code, feel free to contact me at mathias (dot) jackermeier (at) cs (dot) ox (dot) ac (dot) uk.

Please cite this work using the following BibTex entry:

@inproceedings{jackermeier2024dual,
    author = {Jackermeier, Mathias and Chen, Jiaoyan and Horrocks, Ian},
    title = {Dual Box Embeddings for the Description Logic {EL}++},
    year = {2024},
    doi = {10.1145/3589334.3645648},
    booktitle = {Proceedings of the ACM Web Conference 2024},
    series = {WWW '24}
}