Deep Embedded Clustering and Statistical Validation for Student Modality Profiling

Aim/Purpose: This study develops a theory-informed, data-driven framework for identifying student learning modalities using Deep Embedded Clustering (DEC), with the goal of improving scalable and interpretable learner profiling in secondary education. Background: Traditional learning-style identification relies on qualitative or rule-based categorization that may not adequately capture complex learner behaviours. Although deep learning enables richer pattern discovery, many approaches lack interpretability for practical educational use. DEC offers a hybrid alternative by combining representation learning with unsupervised clustering. Methodology: A 48-item questionnaire grounded in the Felder–Silverman Learning Style Model (FSLSM) was administered to senior high school students. After preprocessing and controlled, distribution-preserving augmentation, an autoencoder learned compact latent representations. Clustering was then applied to derive learner profiles, followed by statistical validation and comparative classification experiments to assess structural and predictive utility. Contribution: This study proposes an integrated deep clustering and statistical validation pipeline that produces pedagogically interpretable learner profiles without relying on predefined labels. It extends prior work by demonstrating how unsupervised representation learning can reveal theory-consistent learning patterns. Findings: Four interpretable student profiles emerged, aligning with FSLSM dimensions and differentiated primarily by multivariate response patterns rather than isolated mean differences. Although internal validity indices indicated moderate separation, models trained on latent representations consistently outperformed those trained on raw features, demonstrating the discriminative value of learned embeddings. Recommendations for Practitioners: The identified profiles can support differentiated instruction, including structured visual materials, interactive learning tasks, blended strategies, and targeted scaffolding. Such profiling may enhance adaptive learning systems and early intervention strategies. Recommendation for Researchers: Future studies should validate findings using multi-site, non-augmented datasets and examine cluster stability across cohorts. Lightweight DEC architectures and longitudinal evaluations of personalization strategies warrant further exploration. Impact on Society: By enabling scalable, data-driven personalization, this framework may reduce reliance on subjective assessments and contribute to more equitable and responsive learning environments. Future Research: Further work should explore real-time integration within learning management systems, adaptive feedback mechanisms, and ethical considerations surrounding large-scale learner profiling.