Computer Science editorial
Open AccessOA2026
GroupAffect-4: A Multimodal Dataset of Four-Person Collaborative Interaction
A BIDS-inspired corpus of 40 participants in 10 four-person groups, instrumented with physiology, eye-tracking, and audio, spanning four collaborative tasks and fifteen benchmark targets across individual, interpersonal, and group levels.
Meisam Jamshidi Seikavandi; Alice Modica; Anna Obara; Shan Ahmed Shaffi; Fabricio Batista Narcizo; Tanya Ignatenko; Ted Vucurevich; Karim Haddad; Daniel Barratt; Daniel Overholt; Jesper Bunsow Boldt; Paolo Burelli; Andrew Burke Dittberner· 2026· DOI 10.48550/arXiv.2605.19765
The core problem
Existing affective-computing, social-signal-processing, and meeting corpora capture important parts of human interaction, but they rarely support analysis of affect in co-located groups as a coupled individual, interpersonal, and group-level process. The required signals — per-participant physiology, eye movement, audio, self-report, task outcomes, and personality — are usually fragmented across separate dataset traditions. GroupAffect-4 addresses this gap by introducing a multimodal corpus of 40 participants in 10 four-person groups, each completing four ecologically varied collaborative tasks spanning information pooling, negotiation, idea generation, and a public-goods game. The dataset is designed to support analysis of affect as a coupled process across three levels: within-person state, between-person traits, and group dynamics.
Innovation
The dataset covers over 91% of expected physiology windows and 98% of eye-tracking windows, indicating high data completeness. A clear affective manipulation check across the negotiation block confirms strong task validity. The authors define fifteen benchmarkable targets spanning three analysis levels: within-person state, between-person traits, and group dynamics. Leave-one-group-out feasibility baselines are reported, establishing the dataset's evaluative scope. The fifteen targets are distributed across the three levels as follows:
- Within-person state: targets capturing momentary affective and physiological states.
- Between-person traits: targets capturing stable individual differences such as Big-Five personality.
- Group dynamics: targets capturing emergent group-level processes.
The leave-one-group-out baseline procedure can be formalized as:
where is the set of all 10 groups.
Existing affective-computing, social-signal-processing, and meeting corpora capture important parts of human interaction, but they rarely support analysis of affect in co-located groups as a coupled individual, interpersonal, and group-level process. The required signals — per-participant physiology, eye movement, audio, self-report, task outcomes, and personality — are usually fragmented across separate dataset traditions. GroupAffect-4 addresses this gap by introducing a multimodal corpus of 40 participants in 10 four-person groups, each completing four ecologically varied collaborative tasks spanning information pooling, negotiation, idea generation, and a public-goods game. The dataset is designed to support analysis of affect as a coupled process across three levels: within-person state, between-person traits, and group dynamics.
Each participant is instrumented with a wrist-worn physiology sensor, eye-tracking glasses, and a close-talk microphone. Sessions include continuous affect self-reports, post-task questionnaires, task outcomes, and Big-Five personality scores, all time-aligned to a shared clock. The dataset covers over 91% of expected physiology windows and 98% of eye-tracking windows. Task validity is confirmed by a clear affective manipulation check across the negotiation block. The dataset is released with a BIDS-inspired structure, Croissant metadata, a datasheet, per-session quality reports, and open processing scripts. The experimental design can be summarized as follows:
Why it matters
GroupAffect-4 fills a critical gap in affective computing by providing a multimodal dataset that captures affect as a coupled individual, interpersonal, and group-level process. The combination of per-participant physiology, eye movement, audio, self-report, task outcomes, and personality scores enables researchers to study affect in co-located groups with unprecedented depth. The BIDS-inspired structure, Croissant metadata, datasheet, per-session quality reports, and open processing scripts promote reproducibility and reuse. The fifteen benchmark targets and leave-one-group-out baselines provide a clear evaluative framework for future work. The dataset is publicly archived at https://zenodo.org/records/20037847, and code and processing scripts are available at https://github.com/meisamjam/GroupAffect-4. This resource is expected to accelerate research in social signal processing, affective computing, and group dynamics.
Who should read this
CS practitioners and researchers
Opening member content…