Matching the Expert’s Knowledge via a Counterfactual-Based Feature Importance Measure

To be employed in real-world applications, explainable artificial intelligence (XAI) techniques need to provide explanations that are comprehensible to experts and decision-makers with no machine learning (ML) background, thus allowing for the validation of the ML model via their domain knowledge.

To this aim, XAI approaches based on feature importance and counterfactuals can be employed, although both have some limitations: the last provide only local explanations, whereas the first can be very computationally expensive. A less computationally-expensive global feature importance measure can be derived by considering the instances close to the model decision boundary and analyzing how often some minor changes in one feature’s values do affect the classification outcome.

However, the validation of XAI techniques in the literature rarely employs the application domain knowledge due to the burden of formalizing it, e.g., providing some degree of expected importance for each feature. Still, given an ML model, it is difficult to determine whether an XAI technique may inject a bias in the explanation (e.g., overestimating or underestimating the importance of a feature) in the absence of such ground truth.

To address this issue, we test our feature importance approach both with the UCI benchmark datasets and real-world smart manufacturing data characterized by annotations provided by domain experts about the expected importance of each feature. If compared to the state-of-the-art, the employed approach results to be reliable and convenient in terms of computation time, as well as more concordant with the expected importance provided by the domain expert.