Post Date: 5 June 2026

What Shapes Student Learning? A Multi-Scale Combined Education and Computation Approach

Abstract:

Understanding what shapes student learning remains a central question in educational research and for the pursuit of inclusive and equitable quality education under Sustainable Development Goal 4. This thesis revisits that question by examining how learning-related factors operate across macro, meso, and micro scales. It argues that factors should not be treated only as isolated predictors where only average effects matter, but also as interacting configurations, time-varying states, and within-person processes.

Three empirical projects support this argument. First, a statistical-physics-inspired neural-network model is applied to PISA 2018 data from Chinese Taipei, Hong Kong, Germany, the United States, and the United Kingdom to examine interactions among individual and environmental factors associated with mathematics, science, and reading literacy. Second, recurrent neural networks are used to test whether concentration, motivation, perseverance, engagement, and self-initiative can be detected from behavioural and video-derived traces in two real online courses for Hong Kong secondary students. Third, idiographic lagged and spectral time-series analyses are applied to daily diary data from six adult learners to examine dynamic relations among deliberate practice, emotional experience, perceived progress, and performance.

The findings show that home and cultural resources are broadly facilitative across education systems, whereas leisure ICT use is broadly inhibitory, with several factors showing region-specific roles (Project I). Mindset states are partially detectable under authentic online-learning conditions, although not yet at a level suitable for autonomous deployment (Project II). Daily practice, emotional experience, progress, and performance are coupled through person-specific regulatory architectures (Project III). Overall, the thesis demonstrates how well-structured theory-based computational methods can complement established educational research by making interaction, mindset state variation, and temporal dynamics more visible.