ANTICIPATING STUDENT PERFORMANCE USING FEATURE ENGINEERING AND MACHINE LEARNING CLASSIFIERS

Abstract

Student performance prediction is a core task of Educational Data Mining, yet the contribution of data preparation, as opposed to model choice, is rarely isolated. This study quantifies how label encoding, one-hot encoding and feature selection change the accuracy of six classifiers, Decision Tree, k-Nearest Neighbor, Support Vector Machine, Artificial Neural Network, Random Forest and Extra Trees, and of a voting ensemble on the x-API dataset (480 students, 16 attributes), which was obtained from Kaggle. Three settings are compared: raw features with three grade classes, engineered features (72 attributes after one-hot encoding) with three classes, and a binary pass/fail task with 20 selected features. Information gain, gain ratio, the Gini index and Pearson correlation identified student absence days, visited resources and raised hands as the strongest predictors. With raw features, Random Forest reached 79% accuracy and the voting ensemble 82.6%. After feature engineering, Random Forest reached 90% (+11 points) and the Support Vector Machine improved from 43% to 75%. In the binary setting, Random Forest, Extra Trees, the neural network and the voting ensemble each reached 98.6%. Compared with previously reported x-API results, including a 2024 study, the engineered models are competitive; because merging grade bands makes the task easier, binary results are reported separately from multi-class results. The findings show that data representation produces larger gains than ensembling and that a pass/fail formulation is well suited to early-warning use.

Keywords: Student performance prediction; One-hot encoding; Feature selection; Binary classification; x-API dataset; Random Forest; Educational data mining