[Explainability Evaluation] - Characterization score, and other ways to combine Fid+/- scores

[BlazStojanovic]

🚀 The feature, motivation and pitch

Implement functionality to combine Fidelity +/- metrics (#5958) features for the purpose of evaluation explainability. This feature request is a part of broader explainability evaluation metrics (parent issue #5628) which will be crucial for the new revamp of explainability (see roadmap #5520).

Brief overview

Fidelity measures check the explanations for their faithfulness to the model. In [1] there are several ways proposed for combining both fidelity scores, like

The area under the $\mathrm{Fid}+/(1-\text { Fid-) }$ curve - AUC

The characterization score: $\operatorname{charact} = \frac{w_{+}+w_{-}}{\frac{w_{+}}{f i d_{+}}+\frac{w_{-}}{1-f i d_{-}}}$

Implementation checklist

• Implement the characterization score
• Implement the AUC

References

1. GraphFramEx: Towards Systematic Evaluation of Explainability Methods for Graph Neural Networks