Objective measurement of interpretability in supervised classification

Abstract

Interpretability has become a fundamental requirement for Artificial Intelligence (AI) systems, yet there remains no widely accepted approach for measuring it objectively. Existing approaches propose measures that quantify interpretability, but differ substantially in their underlying assumptions. This dissertation investigates the objective measurement of interpretability by examining the conceptual foundations of proposed measures. It begins with a systematic review of the literature to establish the current state of interpretability measurement. The review identifies ninety peer-reviewed articles proposing computational measures of interpretability, revealing that existing approaches predominantly quantify aspects such as model complexity, robustness, feature interactions, and explanation quality rather than interpretability directly. Building upon these findings, representative computational measures are evaluated against literature-derived consensual notions regarding the relative interpretability of different AI methods. The empirical evaluation demonstrates that although several measures exhibit partial agreement with these notions, no existing approach consistently reflects the expected ordering of interpretability across AI methods. Furthermore, post-hoc approaches are shown to primarily evaluate the quality of generated explanations rather than the inherent information contained within the underlying AI models. The dissertation concludes by investigating whether the consensual notions established from the literature are reflected in human perception through a controlled user study. The findings demonstrate substantial agreement between participant perceptions and the literature-derived notions, providing empirical evidence that these commonly accepted relationships are consistent with human perceptions of interpretability. Collectively, the dissertation establishes a unified framework linking computational measures, literature-derived consensual notions, and human perception. It demonstrates that existing computational approaches quantify properties associated with interpretability rather than interpretability itself, identifies computational measures exhibiting the varying agreement with both literature-derived notions and human perceptions, and provides a principled conceptual and methodological foundation for the development and evaluation of future objective measures of interpretability.