Machine Learning for Student Performance Prediction: A Review of Algorithms, Evaluation Metrics, and Explainability

Abstract

Background: The digitalization of higher education and learning management platforms has generated massive quantities of educational interaction data. Predicting student performance prior to summative assessments represents a cornerstone of contemporary Educational Data Mining (EDM) and Learning Analytics (LA). Objective: This systematic review synthesizes recent developments in machine learning (ML) paradigms for predicting academic trajectories, categorizing modeling paradigms, evaluation frameworks, and post-hoc interpretability methods. Methodology: A multi-perspective analysis was conducted across empirical studies published between 2018 and 2026, evaluating supervised classifiers, ensemble pipelines, deep architectures, and unsupervised clustering frameworks applied to heterogeneous educational datasets. Key Findings: Ensemble methods—specifically Gradient Boosted Decision Trees (XGBoost, LightGBM, CatBoost)—consistently outperform single-model classifiers on tabular academic logs, achieving predictive accuracies exceeding 88–92%. However, performance metrics exhibit severe sensitivity to class imbalance, where uncalibrated models yield artificially high accuracy while failing to identify failing or dropping-out students. The integration of Explainable Artificial Intelligence (XAI) frameworks, predominantly SHapley Additive exPlanations (SHAP) and Local Interpretable Model-agnostic Explanations (LIME), bridges the gap between predictive capacity and pedagogical utility by uncovering non-linear feature interactions such as learning management system (LMS) engagement decay. Implications: While predictive analytics offer substantial potential for early academic interventions, institutional deployment requires calibrated evaluation protocols, strict alignment with data privacy regulations (FERPA/GDPR), and algorithmic fairness audits to mitigate systemic demographic bias.