Multimodal deep clustering of biobehavioral and environmental factors in children’s social skill development using a self-attentive adversarial network

This study proposes a multimodal deep learning framework for modeling the interplay between biological, biomechanical, and environmental factors in children’s social skill development. Using the SAADSC (Self-Attentive Adversarial Deep Subspace Clustering) model, we integrate neurobiological data (e.g., cortisol levels, heart rate variability), biomechanical indicators (e.g., postural control, gait symmetry), and family function variables (e.g., cohesion, stress) to uncover latent clusters reflecting developmental profiles. The model leverages self-attention mechanisms, adversarial training, and residual modules to extract meaningful representations from high-dimensional, heterogeneous data. Five benchmark datasets and real-world social-behavioral data were used to validate the model’s accuracy and robustness, with performance evaluated via ACC and NMI. Results show that our model outperforms traditional clustering and GAN-based baselines, with notable gains from the self-representation and attention modules. The findings support the feasibility of integrating neurophysiological, behavioral, and contextual data through explainable deep clustering, offering novel insights for early identification of social developmental risk and personalized intervention design. While standard benchmark datasets such as MNIST, Fashion-MNIST, Yale B, COIL-20, and USPS were employed to validate the technical stability and robustness of the proposed clustering framework, we also integrated real-world multimodal behavioral, biological, and environmental data to demonstrate the model’s practical relevance for children’s social skill development.