Presentation
Beyond Dyads: Machine Learning for Team Performance Prediction in Multi-Member Teams
DescriptionDynamic team-performance prediction from multimodal psychophysiological data is challenging because performance-relevant patterns are temporally evolving, distributed across heterogeneous sensing modalities, and shaped by higher-order coordination among team members. Existing approaches often rely on static aggregation, shallow multimodal fusion, or individual-level physiological descriptors, which limits their ability to capture team-level dynamics in realistic operational settings. This paper presents MEMDA, a Multimodal Entropy/MI-enhanced Dual-Attention framework for ordinal team-performance prediction. MEMDA represents cardiac, respiratory, neurophysiological, and information-theoretic team-dynamics features as structured temporal sequences. Modality-specific temporal encoders capture within-modality dynamics, intra-modality attention identifies salient temporal regions, and cross-modal attention integrates complementary information across physiological and entropy/average-mutual-information (MI) representations. Experiments on Fire Support Team training data show that MEMDA outperforms representative recurrent and Transformer-based sequence models across temporal window settings, with the DANN-enhanced configuration achieving the strongest average weighted accuracy and weighted F1. Ablation analyses further show that the entropy/MI-enhanced representation is the most informative modality group: removing it consistently degrades performance, while using it alone yields the best single-modality results despite its compact dimensionality. These findings demonstrate that combining multimodal temporal learning with information-theoretic team-dynamics representation provides an effective framework for predicting team performance from realistic, structured, and imbalanced psychophysiological data.
Contributors
Event Type
Lecture
TimeThursday, October 22nd11am - 11:20am PDT
Location
Similar Presentations