A Comparative Evaluation of Transfer Learning Methods for Cross-Context Behavioral Generalization Assessment in Autism Spectrum Disorder Interventions

Abstract

This study presents a comparative evaluation of three transfer learning paradigms—domain adaptation, few-shot learning, and multi-task learning—for assessing cross-context behavioral generalization in autism spectrum disorder (ASD) interventions. Behavioral skill generalization across distinct intervention environments remains a persistent barrier to long-term ASD intervention effectiveness. The distributional discrepancies inherent in behavioral data collected from clinical, school, and home settings create domain shift challenges that standard recognition models cannot adequately address. Drawing primarily on the Self-Stimulatory Behavior Dataset (SSBD) for the video-based comparative analysis, while using the Autism Brain Imaging Data Exchange (ABIDE) as an auxiliary multi-site domain-shift reference and the Expanded Stereotype Behavior Dataset (ESBD) as a published benchmark, this paper establishes an evaluation framework with four quantitative metrics: cross-context recognition accuracy, generalization stability index, few-shot adaptation efficiency, and behavioral transfer success rate. Within this benchmark-informed framework, domain adaptation yields the highest source-proximal accuracy, few-shot learning shows the strongest low-label adaptation, and multi-task learning exhibits the most stable cross-context profiles. These findings provide practical guidance for selecting computational tools to support more scalable and objective ASD intervention evaluation.