In naturalistic group interactions, physiological streams are abundant while affect self‑reports are sparse. Using the GroupAffect-4 dataset (four‑person collaboration, wearable physiology, eye‑tracking, Big Five personality, and post‑task VAD labels), we investigate pseudo‑label augmentation under sparse supervision. Four pipelines are compared: no augmentation, Gaussian Process pseudo‑labelling, personality‑aware trust weighting, and a joint personality‑plus‑confidence weighting, all sharing a common target‑construction framework. Results indicate that, in the known‑team scenario, pseudo‑labels improve over the labelled‑only baseline. The narrow range of Big Five cosine similarities (0.91‑0.99) renders fine‑grained personality weighting ineffective; similarity acts more as a same‑team filter than a calibrated trust signal. With smoothing, augmented SVM variants tie on Valence and Arousal, while the joint personality‑plus‑confidence variant achieves the highest Dominance score. Cross‑subject LOSO transfer remains promising, especially for Arousal, whereas strict session‑isolated LOGO eliminates the augmentation benefit. Given only ten groups, LOGO should be interpreted as a conservative lower bound for unseen‑group transfer. Overall, pseudo‑label augmentation can better exploit sparsely labelled collaborative affect data, while personality information is most useful as an intra‑team selection mechanism.
Review: Pseudo‑labeling shows clear gains for affect sensing in small teams, yet personality similarity offers limited nuance due to homogeneous traits; future work should seek richer individual descriptors or robust cross‑team transfer techniques.