Improving Rare Cheating Detection in Educational Assessments Through Psychometric Feature Engineering and Synthetic Oversampling

The rapid expansion of digital assessment has intensified concerns regarding academic integrity and the detection of dishonest behaviors in technology-enhanced learning environments. Rare cheating detection remains methodologically challenging because verified misconduct labels are difficult to obtain ethically and operationally, while cheating cases represent a small minority of assessment records. This study developed and evaluated a psychometric-informed machine learning framework using an empirical-parameterized synthetic data design. Real assessment data were first used to derive empirical distributions of latent ability, response consistency, item response behavior, and response-time patterns. These parameters were then used to generate a synthetic assessment dataset comprising 1,200 examinee records and 20 items, while cheating-related aberrance parameters were specified using prior evidence from person-fit, aberrant responding, and assessment-security research. The framework integrated latent ability estimates, person-fit statistics, anomaly scores, and response-time indicators within Logistic Regression, Random Forest, and Extreme Gradient Boosting models, with Synthetic Minority Oversampling Technique applied to address class imbalance. Results showed that psychometric feature engineering provided the strongest contribution to rare cheating detection. Person-fit statistics emerged as the most influential predictor, followed by anomaly scores, whereas response-time indicators contributed less when used alone. Synthetic oversampling improved minority-class sensitivity, especially for ensemble models, but its contribution was smaller than that of psychometric feature engineering. These findings support the value of integrating empirical psychometric parameters, theory-informed feature engineering, and interpretable machine learning for scalable academic integrity monitoring.