Gelareh Mohammadi & Patrik Vuilleumier (2022), IEEE Transactions on Affective Computing 13(3):1127–1139 — DOI-10.1109/TAFFC.2020.3028109
Key takeaways
- Data (§III). 139 clips from the emotion-induction literature were piloted, and 99 retained. Workers on CrowdFlower first rated their own personality (BFI-10), then watched a clip and rated 39 CoreGRID items across the five components and 14 discrete emotions (10 kept) on a 5-point scale. That gave 1,567 surveys from 617 workers, one rating per whole clip.
- Structure (§IV.A–B). A dendrogram of emotion profiles splits positive from negative emotions, then gives five clusters (happiness, serenity, distress, anger, surprise). “differences between various emotions can be captured by a few (at least 6) latent dimensions” (p. 1): action tendency, pleasant feeling, novelty, valuation of norms, arousal and goal relevance.
- Prediction (§IV.C). Ordinal regression from all components beats every single component for every emotion, and “no single component is enough to capture the difference of all categories” (p. 8). Without the feeling items, prediction drops most for calm, sadness and satisfaction.
- Personality (§IV.D). Neuroticism correlates with feeling bad and with all negative emotions. Openness and conscientiousness correlate negatively with most discrete emotions, and more strongly with componential items, so “the role of personality on discrete emotion differences can be better justified using the component model” (p. 1).
- Limits (§V). Ratings are global per clip, induction is passive (“the findings of this study, at best, applies to passive emotional experience”, p. 12), and physiology and expression are self-reported.
Standing in the record
Filed on 2026-10-23 because the owner asked whether this Geneva-related work gives a per-episode measure of component synchronization (QUESTION-268). It does not. The word “synchronization” does not occur in it, and each clip yields one retrospective rating vector. See NOTE-1118. Its value to the record is as the behavioural pilot for the fMRI study (LIT-1444), whose design repairs two of its limits: ratings per 12-second event, and recorded heart and respiration rates. It cites Meuleman and Rudrauf (LIT-1440) as the only prior study with the same aim.
It is emotion recognition in affective computing. It reports models but gives no instruction for machine-learning practice.