Predicting Academic Performance Based on Optical Mark Recognition Patterns: A Comparative Study of Transfer Learning Models
Year: 2026
Type: article
Abstract: Optical forms, which are widely used assessment tools in education, have the potential to reflect students' cognitive, sensory, and behavioral characteristics. This study aims to go beyond the traditional use of Optical Mark Recognition (OMR) systems by examining students’ marking behaviors as a predictor of academic achievement. Using 2,100 marking images collected from 42 participants (21 per class), we evaluated 18 transfer learning–based feature extractors, of which 17 were successfully implemented, in combination with 25 classification algorithms. To eliminate potential data leakage and ensure generalization to unseen individuals, all experiments were conducted using 10-fold GroupKFold cross-validation with subject-wise splitting, such that all samples from the same participant were kept within the same fold. The best-performing configuration, EfficientNet-B0 feature representations combined with Support Vector Classification, achieved 88.90% accuracy, with strong threshold-independent performance (ROC-AUC = 0.9268; PR-AUC = 0.8927; Average Precision = 0.8934). Statistical validation via the Friedman test (χ²(16) = 174.34, p < .001) confirmed significant performance differences across transfer learning architectures. These findings indicate that markings on optical forms should not be treated as random artifacts but as behavioral traces that reflect underlying cognitive and affective processes, and they support a shift from a results-oriented to a process-oriented assessment paradigm. From a learning analytics and educational policy perspective, the proposed approach positions paper-based OMR sheets as low-cost “behavior sensors” that can complement early warning mechanisms by enabling earlier identification of students at academic risk and facilitating timely, targeted interventions.
Institutions: Bandırma Onyedi Eylül University, Sakarya University
Language: Turkish
Subfield: Computer Science Applications
Field: Computer Science
Domain: Physical Sciences
SDG: Quality education
Open Access status: diamond
OpenAlex ID: W7142609641