Modified Dwarf Mongoose Optimization for Feature Selection in Imbalanced Student Performance Prediction Tasks

Abstract

Student performance prediction through Educational Data Mining (EDM) methods has become increasingly critical to educational decision-making and intervention. But educational datasets are high-dimensional and imbalanced, presenting serious problems for standard machine learning models. This paper presents an innovative feature selection methodology based on the Modified Dwarf Mongoose Optimization (MDMO), an enhanced version of standard DMO by adding three essential components: adaptive alpha guidance, scout-based diversity, and enhanced babysitter exchange criteria. These modifications boost the exploration-exploitation balance and prevent premature convergence, enabling more efficient search in high-dimensional binary feature spaces. The proposed MDMO is integrated as a wrapper method with five popular classifiers, LogitBoost, linear discriminant analysis, naive bayes, k-nearest neighbors, and decision trees, to form a robust predictive model for student performance. Furthermore, adaptive synthetic oversampling handles data imbalance. Based on real-world student datasets, experimental results indicate that MDMO outperforms other metaheuristic solutions with high classification accuracy and optimal feature reduction. This research illustrates the potential of behavior-inspired role-based swarm intelligence for tackling knotty optimization problems in education.