Comparison and analysis of Türkiye's PISA 2015 science achievement scores based on the statistical regional units classification using machine learning methods

This study aims to predict the PISA 2015 science achievement scores of Türkiye's statistical regions using various machine learning methods, analyze the variables affecting achievement, and examine the performance disparities between regions. The research considers Türkiye's 12 statistical regions, and four machine learning algorithms Random Forest (RF), Support Vector Regression (SVR), Bayesian Regularized Neural Networks (BRNN), and Extreme Gradient Boosting (XGBoost) were employed to predict the science achievement scores for each region. The performance of these models was compared based on their prediction accuracy on test data. Additionally, the importance rankings of the variables were determined using the SHAP (SHapley Additive exPlanations) method, while regional differences were analyzed using the Jaccard similarity index. The findings of the study revealed that the XGBoost and SVR models exhibited the highest prediction performance on the test data compared to other algorithms, while the BRNN and Random Forest models achieved moderate prediction accuracy. These results demonstrate that machine learning algorithms can serve as effective tools for analyzing regional educational data. According to the variable importance rankings performed separately for each region using the SHAP method, the most influential variables on science achievement scores were generally identified as ANXTEST (Test Anxiety Score), ST013Q01TA (Number of Books at Home), ISCEDO (Highest Parental Education Level), BSMJ (Science-Related Career Aspiration), and ST092Q02TA (Educational Resources at Home). The similarity analysis among regions indicated a moderate level of similarity across Türkiye's statistical regions. However, significant disparities in science achievement scores were observed between regions such as TR1 (Istanbul) and TR6 (Mediterranean). Furthermore, differences in the importance rankings of variables were detected in regions like TR2 (Western Marmara) and TR5 (Western Anatolia).