Explainable machine learning analysis of academic resilience among disadvantaged students in Türkiye using Programme for International Student Assessment 2022 data

Academic resilience is a context-specific pattern of positive academic adaptation among students facing socioeconomic disadvantage. This study examined academic resilience in Türkiye using 2022 data from the Programme for International Student Assessment (PISA). Students in the weighted bottom quartile of the national Index of Economic, Social and Cultural Status distribution formed the analytic sample ( n = 1885). For each of the ten PISA mathematics plausible values (PVs), students reaching the corresponding weighted national top quartile were classified as resilient, yielding 204 to 230 resilient students across PV-specific replications. The survey-aware workflow used student weights, school-grouped validation, training-only preprocessing, and minority-class threshold tuning. Weighted logistic regression, Radial Basis Function Support Vector Machine (RBF-SVM), ExtraTrees, Extreme Gradient Boosting (XGBoost), and CatBoost were benchmarked. Although the primary PV-aware comparison favoured CatBoost on Precision-Recall Area Under the Curve (PR-AUC), a three-repeat school-grouped sensitivity analysis found the highest mean PR-AUC for ExtraTrees (0.332, 95% confidence interval [CI] 0.318 to 0.348); logistic regression obtained 0.302 (95% CI 0.241 to 0.340) and CatBoost 0.296 (95% CI 0.267 to 0.316). CatBoost was retained for transparent SHapley Additive exPlanations (SHAP) rather than as a uniquely optimal classifier. The results indicate a limited but detectable group-level signal. Information and Communication Technology (ICT) resources, digital self-efficacy, subject-related ICT use, mathematics anxiety, gender patterning, and digital-emotional interactions jointly characterized the modelled resilience profile. These associations are descriptive and non-causal.