A Robust Student Performance Forecasting Framework Based on Autoencoder and Belief Network Learning

Abstract

Student performance forecasting in the educational sector is of paramount significance because it allows for the early recognition of at-risk students and the ability to give targeted interventions leading to enhanced learning. The majority of the conventional machine learning approaches are hand- and knowledge-engineered feature-based and lack good generalization, while deep models lack good interpretability and require enormous amounts of labeled data. All these disadvantages affect the application of student performance prediction systems on real learning settings. As another approach to such problems, this paper proposes a self-learning system and a hybrid feature selection model that uses the Dingo-Coot Fusion Selector (DiCoot-FS) for feature selection at the level of optimality and the Autoencoder and Belief Network-driven Self-Taught Learning (AutoBel-STL) model for the accuracy of student performance prediction. The highest contribution of this study is its ability to adaptively improve input features, learn informative feature, and apply probabilistic adaptation for improved classification performance. DiCoot-FS integrates dingo-based local exploitation with coot-based global exploration to eliminate redundant features and preserve key academic and behavior variables that affect the performance of students. Concurrently, AutoBel-STL integrates feature compression using autoencoders and probabilistic learning helped by belief networks for self-learning-based adaptation over distributions of unlabeled data. It substantially improves robust predictability and generality on diverse student populations. The proposed framework offers a number of advantages compared to existing solutions, such as improved accuracy, reduced complexity of computation, as well as improved flexibility on handling diversified sets of students. Furthermore, the proposed system performs optimally with the lowest computational time of 1.8 seconds. The proposed method provides an opportunity for top-level educational analysis, enabling colleges to implement individual learning strategies and maximize academic performance for students by measuring and intervening in an initial performance stage.