Understanding faithfulness and monotonicity

Faithfulness in XAI refers to evaluating the correlation between feature importance scores to the actual individual feature’s performance effect on a correct prediction. Measuring faithfulness is typically done by removing pre-determined important features incrementally, observing changes in model performance, and validating feature relevance to a model’s prediction. In other words, are the identified important features genuinely relevant to the final model output?

Besides feature importance correlation, researchers identified additional properties such as polarity consistency for evaluating the faithfulness of explanations, https://doi.org/10.48550/arXiv.2201.12114. Polarity in ML refers to positive and negative analysis – for example, the amount of positive and negative phrases for sentiment analysis. Polarity consistency validates faithfulness by measuring explanation weight based on their contribution and suppression...