Exploring Evaluation Methods for Interpretable Machine Learning: A Survey

Abstract

In recent times, the progress of machine learning has facilitated the development of decision support systems that exhibit predictive accuracy, surpassing human capabilities in certain scenarios. However, this improvement has come at the cost of increased model complexity, rendering them black-box models that obscure their internal logic from users. These black boxes are primarily designed to optimize predictive accuracy, limiting their applicability in critical domains such as medicine, law, and finance, where both accuracy and interpretability are crucial factors for model acceptance. Despite the growing body of research on interpretability, there remains a significant dearth of evaluation methods for the proposed approaches. This survey aims to shed light on various evaluation methods employed in interpreting models. Two primary procedures are prevalent in the literature: qualitative and quantitative evaluations. Qualitative evaluations rely on human assessments, while quantitative evaluations utilize computational metrics. Human evaluation commonly manifests as either researcher intuition or well-designed experiments. However, this approach is susceptible to human biases and fatigue and cannot adequately compare two models. Consequently, there has been a recent decline in the use of human evaluation, with computational metrics gaining prominence as a more rigorous method for comparing and assessing different approaches. These metrics are designed to serve specific goals, such as fidelity, comprehensibility, or stability. The existing metrics often face challenges when scaling or being applied to different types of model outputs and alternative approaches. Another important factor that needs to be addressed is that while evaluating interpretability methods, their results may not always be entirely accurate. For instance, relying on the drop in probability to assess fidelity can be problematic, particularly when facing the challenge of out-of-distribution data. Furthermore, a fundamental challenge in the interpretability domain is the lack of consensus regarding its definition and requirements. This issue is compounded in the evaluation process and becomes particularly apparent when assessing comprehensibility.