Message Passing Query Embedding

This repository contains the implementation of MPQE, proposed in our paper Message Passing Query Embedding. MPQE is a model for answering complex queries over knowledge graphs, that learns embeddings of entities in the knowledge graph, and embeddings for variable types:

Requirements

We recommend creating a virtual environment, e.g. with conda. Use requirements.txt to install the dependencies:

conda create -n mpqe python=3.7
conda activate mpqe
pip install -r requirements.txt

As an additional dependency, we require PyTorch Geometric, which can be installed following these instructions.

Datasets

Datasets can be downloaded from the following links:

• AIFB
• MUTAG
• AM

After downloading them, use tar to decompress them, e.g.

tar -xzf AIFB.tgz

This extracts a folder named AIFB , with two sub-folders: raw contains the triples that make up the Knowledge Graph, plus additional information. The processed folder contains training, validation, and test splits for the query answering task.

Training

From the root directory of the repository, use the following commands to reproduce our experiments (here we show an example when the dataset of queries is at AIFB/processed):

MPQE-TM

python -m mpqe.train --data_dir='AIFB/processed' --adaptive --cuda

MPQE-sum

python -m mpqe.train --data_dir='AIFB/processed' --readout='sum' --cuda

For MPQE-max and MPQE-CMPL, replace 'sum' in the previous command with 'max' and 'concat', respectively.

Using a GTX 1080 Ti NVIDIA GPU, training the models takes around 3 to 6 hours.

Citation

Please use the following entry to cite our work:

@inproceedings{daza2020message,
    title={Message Passing Query Embedding},
    url = {https://arxiv.org/abs/2002.02406},
    booktitle = {{ICML Workshop - Graph Representation Learning and Beyond}},
    author={Daza, Daniel and Cochez, Michael},
    year = {2020},
    arxiv = {2002.02406}
}