Inherently Interpretable Machine Learning: A Contrasting Paradigm to Post-hoc Explainable AI
Patrick Zschech, Sven Weinzierl, Mathias Kraus · Business & Information Systems Engineering · 2025
A decade ago, data scientists devoted roughly 60–80% of their effort to manual feature engineering, aiming to create features with high predictive power and meaningful properties (Press 2016 ). However, with the rise of deep learning and highly flexible models capable of processing raw input features and learning intricate feature representations, this focus has shifted dramatically (Janiesch et al. 2021 ). Today, data scientists spend much of their time explaining opaque machine learning (ML) models because users struggle to understand their complex decision logic (Bauer et al. 2021 ).