Improving Predictions of Student Performance and Personalized School Guidance with Hybrid Machine Learning Models

Fighting against school failure is one of the major challenges of education systems today. For decades, several measures have been put in place in education systems in order to provide solutions to this scourge, in particular, new pedagogic, online guidance platforms, the use of a qualified teacher, a school coach, specialized classes, and tutoring. However, these solutions have given rise to problems in the connection between the individual and the collective, which manifests itself by the difficulty for the teaching teams to take into account the specificity of each student, without weakening the collective. The digitization of the education system nowadays offers the possibility of exploiting data produced in very large quantities throughout the teaching–learning process to specifically follow each student. The results of this exploitation for previous classes should be made available to academicians in order to help them fight against school failure. In this work, based on educational data mining (EDM), we propose a hybrid approach based on the K-Nearest Neighbors (KNN) algorithm, multiple linear regression, and decision tree (ID3 version) that enables academicians to predict student's grade at the end of academic year based on his general average and his disciplinary behavior of previous classes and propose a solution to improve their level. Our model was developed in two phases: training on labeled data to identify trends, then validation on unlabeled data to test its generalization. Validating our approach with data from two secondary schools in Cameroon produces a prediction that is 60% for KNN algorithm and 52.72% for multiple linear regression reliable compared to actual data and allows us to conclude that factors such as age, gender, and behavior disciplinary are elements that greatly influence a student's work.