Method Agnostic Model Class Reliance (MAMCR) Explanation of Multiple Machine Learning Models

Various Explainable Artificial Intelligence (XAI) methods provide insight into the machine learning models by quantitatively analysing the contribution of each variable to the model’s predictions globally or locally. The contribution of variables identified as (un)important by one method’s explanation may not be identified as the same by another method’s explanation for the same machine learning (ML) model. Similarly, the important feature of many well performing ML models that fit equally well on the same data (which are termed as Rashomon set models) may not be the same as each other. While this is the case, providing the explanation based on a single model in the lens of a specific explanation method would be biased over the model/method. Hence, a framework is proposed to describe the consensus variable importance across multiple explanation methods for many almost-equally-accurate models as a method agnostic explanation for the model class reliance. Empirical experiments are carried out on the COMPAS dataset with six XAI (the Sage, Lofo, Shap, Skater, Dalex and iAdditive) methods for verifying whether an inadmissible feature becoming an (un)important feature is consistent across multiple explanation methods and getting the consensus explanation. The results demonstrate the efficiency of the method agnostic model class reliance explanation and its coverage to the model reliance range of all the almost-equally-accurate models of the model class.