Pseudo-Label Augmentation for Affect Sensing in Small Collaborative Groups
- ,
- Tanya Ignatenko,
- Fabricio Batista Narcizo,
- ,
- Jesper Bünsow Boldt,
- Andrew Burke Dittberner
- ,
- ,
- ,
- GN Store Nord A/S,
- ,
Research Output:
Conference Article in Proceeding or Book/Report chapter
Article in proceedings
Research
Peer-reviewOpen access
Abstract
Physiological affect sensing in naturalistic group interaction is typically limited more by label density than by sensor density: wearable devices produce thousands of time windows, while self-report labels are collected only a few times per session. This paper asks how to mine more supervisory signal from sparsely labelled group sessions. More broadly, we treat this as a step toward collective-state sensing: better individual affect estimates within a team help reveal the affective structure emerging at group level. Using GroupAffect-4, a four-person collaborative dataset with wearable physiology, eye tracking, Big Five personality, and post-task VAD labels, we study a progression of label-mining strategies: no augmentation, Gaussian Process (GP) pseudo-labelling, personality-aware trust weighting, and joint personality-plus-confidence weighting under a shared target-construction pipeline. We further ask whether individual predispositions (here, Big Five personality) together with group-level context improve which pseudo-labels should be trusted. They help, but not in the originally hypothesised graded way: the compressed BFI cosine-similarity range (0.91–0.99) renders weight ordering negligible (perm p = 0.39), so personality similarity acts as a same-team membership filter rather than a calibrated trust dial. In the known-team setting, pseudo-label augmentation improves over the labelled-only baseline, with the clearest benefit coming from within-team filtering rather than from fine-grained personality ranking. With smoothing, A1–A3 are tied to rounding on Valence and Arousal, while A3 has the highest Dominance score among the augmented SVM variants; the practical choice is therefore dimension-dependent. Cross-subject LOSO transfer remains encouraging for such a small and sparsely labelled dataset, especially on Arousal, whereas strict session-isolated LOGO removes the augmentation benefit. At the current scale of only 10 groups, LOGO should be read as a conservative lower bound on unseen-group transfer rather than as a definitive ceiling, but the present evidence shows that personality weighting mainly operates as a within-team mechanism.
Publication Information
Output type
Research Output:
Conference Article in Proceeding or Book/Report chapter
Article in proceedings
Research
Peer-reviewOriginal language
EnglishPublication milestones
- Accepted/In press - 2026
- Published - 02/10/2026
Publication status
Published - 02/10/2026
Publisher
Association for Computing Machinery, United StatesHost publication title
Companion of the INTERNATIONAL CONFERENCE ON MULTIMODAL INTERACTION (ICMI Companion '26Access to documents
Related Event
Title
28th International Conference on Multimodal Interaction
Event type
ConferenceDegree of recognition
International eventDate
05/10/2026 - 09/10/2026Location
NapoliItaly
