This paper investigates the suitability of machine learning methods for discovering ground truth about psychological categories, using emotion as a test case. It compares supervised classification (using emotion labels) with unsupervised clustering (without labels) across three datasets measuring brain, body, and subjective experience during emotional episodes. The study finds that supervised and unsupervised methods yield inconsistent results, suggesting caution when interpreting machine learning analyses to test psychological hypotheses. The authors propose that either emotion category labels accurately reflect biological categories but measurement limitations hinder discovery, or that folk emotion categories are not sufficiently useful for organizing biological signals.

Bahar Azari, Christiana Westlin, Ajay B. Satpute, J. Benjamin Hutchinson, Philip A. Kragel, Katie Hoemann, Zulqarnain Khan, Jolie B. Wormwood, Karen S. Quigley, Deniz Erdogmus, Jennifer Dy, Dana H. Brooks, Lisa Feldman Barrett