Benchmarking and tuning regression algorithms on predicting students’academic performance

With the adaption of online learning environment, students’ learning behavior can be recorded as digital data. In order to implement the conceptual framework of learning analytics, many researchers applied machine learning methodologies and used data which collected from digital learning environment to predict students’ academic performance for targeting at-risk population. However, along with the characteristic of machine learning methodologies, it presents diversity prediction performance due to the statistical property of educational data and these caused the difficulty to applied machine learning technology to classroom. In this study, we collected the state-of-the-art on regression algorithms and used an E-book-based learning dataset within 53 students for benchmarking the suitable algorithm for targeting at-risk students. In addition, we address the issues from learning environment, including over-concentration score, dropout students and data instance insufficiently, for improving prediction performance. The results revealed that the proposed performance tuning process could obtain optimal performance metrics and avoid over-fitting problem.