A time series data-driven students’ academic performance modelling using Multivariate Adaptive Regression Splines

Abstract

This work presents a data-driven approach to predictive modelling of student academic performance using Timeseries and Multivariate Adaptive Regression Splines (MARS), a flexible machine learning technique proficient at handling the non-linear and interactive effects present in the student academic performance dataset. The time series Autoregressive Integrated Moving Average (ARIMA) models the linear time-based variation of the student’s data. On the other hand, MARS, provides a non-parametric, piecewise linear modelling framework that automatically identifies important predictors and interactions for the complex relationships found in academic datasets. The research utilized the academic records of 250 computer science students spanning five cohorts from 2014 to 2018. Key predictors included demographic data and course-specific grades, with Final Cumulative Grade Point Average (FCGPA) as the target variable. The ARIMA-MARS model, implemented in R and optimized using the nprune value for parameter tuning, identified 16 out of 76 predictors based on hinge functions, which highlight critical performance indicators based on grades and interactions. The model achieved a Root Mean Squared Error (RMSE) of 0.201 and Mean Absolute Error (MAE) of 0.159, indicating strong predictive accuracy and generalization capability. This approach supports data-driven, targeted interventions to enhance academic performance. Future research may expand the cohort size or explore alternative hybridized machine learning models for comparative analysis and broader applicability.