PREDICTING SCHOOL NUMERACY PERFORMANCE THROUGH EDUCATIONAL DATA MINING: AN INTEGRATED K-NEAREST NEIGHBOR AND MULTINOMIAL LOGISTIC REGRESSION APPROACH

Abstract

International large-scale assessments such as PISA consistently indicate that numeracy achievement in Indonesia remains below the international average, highlighting the need for early identification of schools requiring targeted educational support. Although previous studies have examined factors associated with numeracy performance, few have integrated predictive machine learning with explanatory statistical modelling to support evidence-based educational decision making. This study aimed to develop a hybrid analytical framework by combining the K-Nearest Neighbor (KNN) algorithm optimized using the Synthetic Minority Over-sampling Technique (SMOTE) with Multinomial Logistic Regression to predict school numeracy performance and identify its significant determinants. Secondary data were obtained from the 2023 Indonesian Education Report, comprising 181 elementary schools in Makassar, South Sulawesi. The predictive model utilized four educational indicators: reading literacy, formulating, employing, and interpreting competencies. The optimized KNN model (k = 5) achieved a macro classification accuracy of 90%, demonstrating strong predictive performance across numeracy proficiency levels. Furthermore, the trained model successfully classified an unseen school profile into the Basic numeracy category, illustrating its practical applicability for early institutional mapping. Multinomial Logistic Regression revealed that all predictor variables significantly influenced numeracy performance (p < 0.05), with the employing competency emerging as the strongest determinant, indicating that higher application skills substantially increased the likelihood of schools progressing from the Basic to the Proficient category. Reading literacy also exhibited a significant positive contribution, reinforcing the close relationship between language comprehension and contextual mathematical problem solving. These findings demonstrate that integrating machine learning prediction with statistical interpretation provides an effective data-driven framework for supporting educational policy, identifying at-risk schools, and designing timely interventions to improve numeracy achievement.