Evidential Relational-Graph Convolutional Networks for Entity Classification in Knowledge Graphs

This repository contains a Pytorch-based implementation to reproduce Evidential Relational-Graph Convolutional Networks for Entity Classification in Knowledge Graphs, as published in CIKM 2021, as well as more general code to leverage evidential learning to train Graph Convolutional Networks to measures uncertainty in knowledge graphs. The code is adpated from Kipf's Keras-based implementation. See details in Modeling Relational Data with Graph Convolutional Networks (2017).

Setup

To use this package, you must install the following dependencies first:

• Compatible with PyTorch 1.4.0 and Python 3.7.3.
• Dependencies can be installed using requirements.txt.

Training

You can learn the representations on the datasets AIFB, MUTAG, BGS and AM using E-RGCN by specifying the dataset with the argument --dataset. This file loads the knowledge graphs and learns a model for node classification. We include early-stopping mechanisms in pytorchtools.py to pick the optimal epoch.

• AIFB:

python run.py --data aifb --epochs 50 --bases 0 --hidden 16 --lr 0.01 --l2 0

• MUTAG:

python run.py --data mutag --epochs 50 --bases 30 --hidden 16 --lr 0.01 --l2 5e-4

• BGS:

python run.py --data bgs --epochs 50 --bases 40 --hidden 16 --lr 0.01 --l2 5e-4

• AM:

python run.py --data am --epochs 50 --bases 40 --hidden 10 --lr 0.01 --l2 5e-4

Note: Results depend on random seed and will vary between re-runs.

• --bases for RGCN basis decomposition
• --data denotes training datasets
• --hidden is the dimension of hidden GCN Layers
• --lr denotes learning rate
• --l2 is the weight decay parameter of L2 regularization
• --drop is the dropout value for training GCN Layers
• Rest of the arguments can be listed using python run.py -h

Results

The following image shows the results reported in the Paper. Considering the results for entity classification we observe on the one hand that our model E-R-GCN achieves state-of-the-art results on AIFB and AM, outperforming R-GCN in particular, and on the other hand that the results reported in R-GCN are confirmed. The experimental results show a performance increase compared to the previous R-GCN model.

The use of the evidential learning approach makes no overconfident predictions and allows to specify uncertainties of the model and thus to accurately determine whether a model is suitable for prediction on a subgraph, which is currently not considered in classical ML approaches. This is particularly important when considering dynamic changes in knowledge graphs or to detect out-of-distribution samples.

Citation

If you use this code for evidential learning as part of your project or paper, please cite the following work:

@inproceedings{weller2021ergcn,
    author = {Weller, Tobias and Paulheim, Heiko},
    title = {Evidential Relational-Graph Convolutional Networks for Entity Classification in Knowledge Graphs},
    year = {2021},
    isbn = {9781450384469},
    publisher = {Association for Computing Machinery},
    address = {New York, NY, USA},
    url = {https://doi.org/10.1145/3459637.3482102},
    doi = {10.1145/3459637.3482102},
    booktitle = {Proceedings of the 30th ACM International Conference on Information & Knowledge Management},
    pages = {3533–3537},
    numpages = {5},
    keywords = {evidential learning, graph convolutional neural network, knowledge graph, entity classification},
    location = {Virtual Event, Queensland, Australia},
    series = {CIKM '21}
}

Licence

Shield:

This work is licensed under a Creative Commons Attribution-ShareAlike 4.0 International License.