Deep Learning-Based Interpretative Analysis of Predictive Variable Structures Across Creative Thinking Domains in PISA 2022

This study developed and applied a deep learning-based prediction model (TabNet) using task response data from the 2022 PISA Creative Thinking assessment. It aimed to analyze structural differences in key predictors across four domains: written expression, visual expression, social problem solving, and scientific problem solving. Moving beyond the limitations of self-report measures, the study utilized actual performance-based data and employed IRT-based scoring and stepwise variable combinations to examine the influence of cognitive achievement, socioeconomic background, educational context, and socio-emotional characteristics. Results showed that cognitive achievement was a strong predictor in all domains, but the structure of influential variables varied significantly across domains. In particular, the visual expression domain demonstrated limited explanatory power with conventional variables. The findings highlight the necessity of domain-specific approaches to creativity prediction and suggest the potential of deep learning models for use in educational assessment and instructional design.