Prediction of Secondary-School Student Cognitive Performance in Informatics

Machine learning techniques for the prediction of performance in the learning process are primarily studied at the post-secondary level of education. Data sets are usually large at that level, resulting in predictive models that have a high accuracy. In contrast, limited research has been conducted at the secondary school level, mostly due to the typically small data sizes and the unique educational challenges of that level. In the present study we aimed to address such issues by implementing a model to predict the final performance of lower secondary school students in a course on Informatics, using a teaching scenario that combined the flipped classroom, a variation of the jigsaw technique and educational robotics. Given the student grades from the first third of the year, as well as demographics data, we used an IBM data analytics tool to create and test several predictive models. The CHAID algorithm achieved the highest accuracy (82.14%) and AUC value, outperforming others like Quest, C5, BN, and RF. The tool also pinpointed the educational activity that was the most significant predictor of final performance, indicating its strong instructional value. Despite the relatively small dataset, the results suggest that careful parameter selection can yield models that predict learner performance with a high accuracy and assist educators in continuously improving their teaching for better student outcomes.