Article: An AI-driven multimodal fusion and graph-temporal modelling framework for developmental assessment of social competence and emotional expression in preschool children Journal: International Journal of Reasoning-based Intelligent Systems (IJRIS) 2026 Vol.18 No.17 pp.15 - 28 Abstract: Current multimodal behaviour analysis focuses on isolated tasks without a unified framework for social behaviour, emotion and developmental assessment. To tackle this issue, we propose a multi-modality fusion and graph-temporal modelling framework, which combines visual, audio, motion, and environmental signals via residual attention, graph convolutions, bidirectional GRU coding and joint behaviour-emotion learning. Experiments show that the proposed framework reaches 0.92 accuracy and 0.88 F1 in the task-specific evaluation setting, while the AUC of emotion recognition reaches 0.91. Although the revised comparative experiments indicate that Transformer-based cross-modal sequence models obtain higher Accuracy and AUC on the current dataset, additional interpretability evaluation demonstrates that the proposed framework achieves higher indicator-alignment correlation, stronger evidence-segment overlap with expert annotations, and higher educator utility ratings in observation-oriented analysis. This framework offers unified pipeline for joint behaviour-emotion modelling and developmental indicator mapping, while delivering practical value for preschool monitoring, fine-grained classroom observation and data-driven educational decision-making. Inderscience Publishers - linking academia, business and industry through research

Title: An AI-driven multimodal fusion and graph-temporal modelling framework for developmental assessment of social competence and emotional expression in preschool children

Authors: Xia Xiao; Shaomei Li

Addresses: Faculty of Education, Changzhi University, Changzhi, 046000, Shanxi, China ' Faculty of Education, Shaanxi Normal University, Xi'an, 710000, Shanxi, China

Abstract: Current multimodal behaviour analysis focuses on isolated tasks without a unified framework for social behaviour, emotion and developmental assessment. To tackle this issue, we propose a multi-modality fusion and graph-temporal modelling framework, which combines visual, audio, motion, and environmental signals via residual attention, graph convolutions, bidirectional GRU coding and joint behaviour-emotion learning. Experiments show that the proposed framework reaches 0.92 accuracy and 0.88 F1 in the task-specific evaluation setting, while the AUC of emotion recognition reaches 0.91. Although the revised comparative experiments indicate that Transformer-based cross-modal sequence models obtain higher Accuracy and AUC on the current dataset, additional interpretability evaluation demonstrates that the proposed framework achieves higher indicator-alignment correlation, stronger evidence-segment overlap with expert annotations, and higher educator utility ratings in observation-oriented analysis. This framework offers unified pipeline for joint behaviour-emotion modelling and developmental indicator mapping, while delivering practical value for preschool monitoring, fine-grained classroom observation and data-driven educational decision-making.

Keywords: early intervention; child development assessment; attention mechanism; group interaction modelling; multimodal behaviour analysis; affective computing; preschool observation; educational decision support.

DOI: 10.1504/IJRIS.2026.154537

International Journal of Reasoning-based Intelligent Systems, 2026 Vol.18 No.17, pp.15 - 28

Received: 20 Jan 2026
Accepted: 29 Apr 2026
Published online: 02 Jul 2026 *

Title: An AI-driven multimodal fusion and graph-temporal modelling framework for developmental assessment of social competence and emotional expression in preschool children

Keep up-to-date