A Cumulative Increasing Kemelized Nearest-Neighbor Bagging Method for Early Course-Level Study Performance Prediction

Abstract

Early course-level study performance prediction is a significant educational data mining task to forecast the success of each current student in a course using the historical data of the students in the previous same course. This task can be resolved by different machine learning approaches in various educational contexts. However, how easily and effectively a solution is deployed in practice is restricted by many factors. Two main factors that have not yet been discussed simultaneously are incremental mining and interpretability when the task is prolonged course after course. Therefore, in this paper, we propose a novel cumulative increasing kernelized nearest-neighbor bagging method for early course-level study performance prediction. Our method is a lazy learning one with an inherent incremental mining mechanism, defined as an ensemble method. Although it works in a feature space to handle a non-linearly separated data space, interpretability is enabled with instance-based learning and a confidence score of each prediction is further provided for practical applications. Experimental results on several public datasets confirm the effectiveness of our method as compared to other traditional prediction methods and well-known ensemble ones. Its better early predictions can help both students and lecturers make appropriate course changes for students’ ultimate success.