Explainability Evaluation Metrics

🚀 Explanability Evaluation

Provide basic evaluation metrics, such as fidelity+/- measures (taxonomic survey, GraphFramEx) to evaluate the quality of explanations.

Explanations for GNNs typically take 2 forms:

• Importance scores / contribution of feature dimensions for nodes / edges of the input graph. The higher the contribution / scores are, the more influence the features have for prediction of a certain label.
• A soft / hard masks that selects a subset of edges that are important for the prediction

There are many dimensions to consider when making these evaluations, including

• Explanation of underlying phenomenon vs. model prediction behavior
• Soft / hard masks as a form of explanation
• Thresholding of masks to produce a concise explanation More details can be found in https://arxiv.org/abs/2206.09677 See more on different problem settings in https://github.com/pyg-team/pytorch_geometric/issues/5629. The difference is that we are considering explanation evaluation here.
• Fidelity + / - metrics
• Characterization score, and other ways to combine Fidelity +/- metrics (e.g. AUC)
• Faithfulness - https://arxiv.org/pdf/2208.09339.pdf We can potentially utilize https://github.com/understandable-machine-intelligence-lab/Quantus if applicable
• AUC, accuracy, precision, recall metrics when groundtruth explanations are available. The metric API could be groundtruth_eval(explanation, groundtruth), where explanation is the Explanation class in #5804 and groundtruth is the groundtruth explanation (e.g. motif structure in #5817 ). May require hard explanation (thresholded) for now. But we can potentially extend to soft explanations in future.

In future, we will also consider evaluations for counterfactual explanations, model-level concept explanation etc. But they are out of the scope for this issue.

Alternatives

No response

Additional context

No response