Why Machine Learning Needs Theoretical Guidance to Support Future Theory Building

Machine learning has long held intuitive appeal in aiding theory building, owing to its capacity to automatically uncover intricate patterns from vast amounts of observed data. Yet, recent literature in organizational research has acknowledged a potential hurdle in utilizing machine learning for theory building: a lack of consistency in the machine-learning output. This means that the same machine learning algorithm could learn predictively accurate yet theoretically contradicting patterns when applied to different samples from the same population, or even when run multiple times on the same sample. This article aims to address the fundamental question of whether, when, and how predictive accuracy implies theoretical pertinence in the context of using machine learning to support inductive theory building. Specifically, we offer theoretical arguments to establish the importance of ensuring that a machine learning algorithm’s assumptions regarding input data are properly aligned with the phenomena being studied. Building on these arguments, we develop a 2×2 framework that outlines the conditions under which four distinct types of machine learning algorithms may be better suited to facilitate inductive theory building in behavioral and organizational research.