WAVELET-KOOPMAN PROTOTYPE TRANSPORT LEARNING FOR CROSS-DATASET CHILD AFFECT AND ENGAGEMENT GENERALIZATION

A significant challenge in automated child affect and engagement analysis is transferring a model across datasets, as recording conditions, child behaviour, temporal effects, and annotation schemes vary and can seriously impact model reliability. To address this challenge and enable robust cross- dataset behavioural representation learning, this paper proposes Wavelet–Koopman Prototype Transport Network (WKPT- Net) and a Bilevel Spectral–Prototype Transport Search (BSPTS) strategy for WKPT- Net. Using a three- band wavelet analysis, the framework decomposes facial and movement characteristics into three temporal layers: slow, intermediate, and fast. Nonlinear behavioural evolution is then modelled by band- wise Koopman latent operators, with complementary temporal information fused via a spectral gate. Prototype memory and Sinkhorn- based optimal transport transfer latent structures between DAiSEE and EmotiW without mapping the non- identical label space to a common classification head. All parameters, including wavelet configuration, latent dimensionality, transport regularization, prototype parameters, and computational complexity, are jointly optimized by BSPTS. For the same deterministic prototype evaluation, WKPT- Net achieved accuracy and macro- F 1 scores of 96. 2% and 95.2. 2%, respectively, on the DAiSEE dataset, whereas the dataset- specific EmotiW emotion head achieved 94. 8% and 95.2. 2%, respectively. There was a 68.9% reduction in cross- domain discrepancy when normalized, while prototype transport achieved separation below 6% of inter- prostate distance, indicating good system alignment without collapsing. The proposed architecture thus shows the promise of multiscale temporal decomposition, Koopman dynamics, prototype learning, and optimal transport in transferable behavioural modelling. The results help underpin the creation of scalable frameworks for AI- assisted engagement monitoring and individualized behaviour assessment with mental- health- oriented perspectives.