Data Generation using a Probabilistic Auto-Regressive Model with Application to Student Exam Performance Analysis

Exam scores are usually the most important assessment criterion for evaluating students' understanding of course materials and learning outcomes. Mid-term scores can serve as an indicator to predict exam scores because these assessments are generally common in the format of questions, closed-book assessments, and covering a variety of course topics. Machine learning algorithms have become a promising educational technology for forecasting student exam scores. If the prediction model suggests that a student is academically at risk, additional tutorial sessions and academic advice are potential follow-up actions. However, there are several research challenges: (i) mid-term scores are small-scale with limited samples, (ii) many physiological signals are required to supplement midterm and final exam scores as a multimodal problem, and (iii) insufficient investigation of the length of physiological signals. This paper employed a probabilistic autoregressive (PAR) model to generate Photoplethysmography (PPG) signals. A Wearable Exam Stress Dataset was used to benchmark the model. Results revealed that the PAR model can synthesize high-quality PPG signals in midterm and final exam scores to support educational research, particularly in the student exam performance analysis. Future research directions are suggested to further explore data generation research in education.