Early prediction of student performance based on behavioral data in blended learning

Purpose Predicting learner outcomes in blended learning (BL) is a new problem with many challenges, as learner data must be collected in both face-to-face and online environments. The purpose of this article is to identify the best method for building a model to predict student performance in BL and to determine the appropriate time for early prediction. Design/methodology/approach In this study, the authors propose a process for building a model to predict students’ learning outcomes early in BL. We will identify important features, select appropriate machine learning algorithms to build prediction models and determine suitable early prediction times. We collected learner behavior data at various times in two courses: general information (GI) and data structures and algorithms (DSA), with 746 and 102 learners, respectively. Findings Experimental results show that the two most suitable algorithms for building prediction models are linear regression and random forest classification algorithms. Additionally, the research reveals that offline classroom behaviors such as attending class diligently and taking quizzes contribute to improving the efficiency of prediction models. Moreover, early prediction can be done from the eighth week of the semester with an efficiency that is not much different from predicting using the whole course data. Originality/value Early prediction of learning outcomes allows educators to adjust teaching plans and warn students, thereby improving the quality of teaching and learning. Although the research was only conducted in small-scale experimental courses, the results are positive and can be applied to other blended courses with the same teaching scenario.