As stated in Wang et al. - Explanation Guided Contrastive Learning for Sequential Recommendation, model fidelity is a good measure of evaluation for explanation models. For what concerns sequential recommendation explanation, which is the task that the paper referenced above tackles, the measure tells us what’s the percentage of the recommendation results the can be explained by the model.
Fidelity in Counterfactual Generation
In other words, it tells us the percentage of points on which the model is able to generate a valid counterfactual.
Example: an 80% fidelity means that for 80% of the points in input to the counterfactual generation model, a valid counterfactual is generated; while the other 20% of points result in a non-valid counterfactual, meaning a different input which has the same outcome when input to the oracle (black box model).
Usually a good fidelity spans between 70% to 90%.
tags: recommender-systems performance-evaluation ai-explainability related: Fidelity-Interpretability trade-off