Student Performance Prediction Using Hybrid Ensemble

This study uses the UCI Student Performance dataset to present a novel weighted hybrid ensemble classifier for predicting academic performance. The proposed model integrates Random Forest (RF), XGBoost (XGB), and a Neural Network (NN), with dynamically adjusted weights based on the influence of prior grades (G1 and G2). Initially set at 40% RF, 30% XGB, and 30% NN, the weights shift to 25% RF, 25% XGB, and 50% NN when G1 and G2 are identified as highly influential, enhancing adaptability to grade-driven patterns. The dataset comprises 395 students and 32 features, with final grades (G3) categorized into Low (0–9), Medium (10–14), and High (15–20). Evaluated on a 20% test split, the model achieves an accuracy of 0.82. Our contributions include the design of a feature-sensitive dynamic ensemble architecture, the development of an interpretable decision-support interface via Streamlit, and the implementation of a prediction framework tailored to real-world academic variability. Experimental results demonstrate that the model effectively balances adaptability and accuracy, with prior grades G1 and G2 emerging as dominant predictors (combined importance: 0.575), significantly improving the classification of student performance levels.