Out-of-Distribution Robust Explainer (ORExplainer)

Requirements

To ensure compatibility, we recommend using the following versions:

• python==3.10.10
• spacy==3.5.2
• torch==1.13.1
• pytorch-geometric==2.2.0

Additional dependencies and exact package versions can be found in the environment.yml file.

Setting Up the Environment

To create the environment with all dependencies:

conda env create -f environment.yml

Usage

The main scripts for running the model on synthetic and real datasets are provided below:

Running the Pipelines

Synthetic Datasets

python pipeline_syn.py --dataset=syn1 --ood=1 --date=0301

• Replace syn1 with your target synthetic dataset (syn1-syn4).
• The --ood flag specifies the OOD Structure-level OOD (0–3).

Real-World Datasets

python pipeline_real.py --dataset=Cora --ood=2        # Feature-level OOD (level 2)
python pipeline_real.py --dataset=Citeseer --ood=label1  # Unseen-label OOD

• Structure-level / Feature-level OOD: set --ood to an integer between 0–3 (higher = stronger OOD level).
• Unseen-Label OOD: set --ood to label0 (disabled) or label1 (enabled).
• dataset: choose from Cora or Citeseer.

Recommended Practice

For better traceability of results, we recommend appending a date-based identifier to your save directory or experiment name.

Hyperparameter Settings

OOD Type	Dataset	Learning rate	Epochs	$\alpha$ (coff_size)	$\beta$ (coff_ent)	$\gamma$ (WEP)
Structure	syn1	0.003	10	0.05	1.0	5.0
	syn2	0.003	20	0.05	1.0	5.0
	syn3	0.003	20	0.1	1.0	5.0
	syn4	0.003	30	1.0	1.0	10.0
Feature	Cora	0.005	20	1.0	$5×10^{-4}$	0.1
	Citeseer	0.005	20	1.0	$5×10^{-4}$	0.1
Unseen	Cora	0.005	20	1.0	$5×10^{-4}$	0.1
	Citeseer	0.005	20	1.0	$5×10^{-4}$	0.05

References

This project is based on methodologies from:

1. DIG Library - A Deep Graph Library that provides a framework for graph learning.
2. RE-Parameterized Explainer - From LarsHoldijk/RE-ParameterizedExplainerForGraphNeuralNetworks, an approach for graph neural network explanations.