Machine Learning Reveals What Helps Disadvantaged Students in Türkiye Beat the Odds in Mathematics

Some students seem to defy the statistics. They grow up in households with few books, limited income, and scarce educational resources, yet they land in the top quarter of mathematics performance in their country. Researchers call this academic resilience, and a new study of Turkish 15-year-olds has used explainable machine learning to ask a deceptively simple question: what actually distinguishes these students from equally disadvantaged peers who fall behind? The answer, drawn from the 2022 wave of the Programme for International Student Assessment (PISA), is not a single magic ingredient but a conditional profile in which digital resources, confidence with technology, and low mathematics anxiety combine in ways that neither factor could achieve alone.

The study, conducted by Mehmethan Erduran and Kadir Demir of İzmir Demokrasi University together with Elif Buğra Kuzu Demir of Dokuz Eylül University, was published in Discover Education. The team started from the full Turkish PISA 2022 sample of 7,250 students and narrowed it to 1,885 disadvantaged students drawn from 179 schools. Disadvantage was defined statistically: students had to fall at or below the weighted first quartile of Türkiye’s distribution on the PISA Index of Economic, Social and Cultural Status, a standard composite measure of family wealth, parental education, and home possessions. A student was labelled resilient only if he or she simultaneously reached the top national quartile in mathematics, a bar that sat at roughly 514 to 516 points depending on the measurement replication used.

That threshold rule matters because resilience is a relative concept. Against an international benchmark, very few Turkish disadvantaged students would qualify, since the country’s average mathematics performance sits below the OECD average, which would blend national performance differences with within-country resilience. Using national quartiles keeps the question firmly within Türkiye’s own opportunity structure. Even so, resilience remained rare: across ten parallel analyses, only 204 to 230 students met the criteria, a weighted prevalence of about 11.7 percent. That figure closely matches the OECD’s own country note, which reports that roughly 12 percent of disadvantaged Turkish students reach the top national quarter of mathematics performance.

Because the outcome was so rare, the researchers took unusual care with their machine learning pipeline. They benchmarked five model families: weighted logistic regression as a traditional statistical baseline, a radial basis function support vector machine, and three tree-based ensembles known as ExtraTrees, XGBoost, and CatBoost. Student survey weights were retained throughout so that the models reflected Turkey’s nationally representative sampling design rather than an unweighted caricature of it. Crucially, because students are nested within schools, all train-test splits were grouped by school rather than by student, preventing the leakage that would otherwise produce inflated performance estimates. Data preprocessing steps, such as median imputation of missing questionnaire responses, were fitted only on training data, and decision thresholds were tuned on validation sets to compensate for the severe class imbalance.

The team also grappled with a quirk of PISA measurement: achievement is reported not as a single score but as ten statistically interchangeable plausible values, each representing one possible draw from the distribution of a student’s likely ability. Rather than averaging them, the researchers ran the entire modelling workflow ten separate times, once per plausible value, and summarized the results. This preserved the genuine uncertainty in the achievement construct and ensured that the final findings were robust to measurement noise rather than artefacts of one particular scoring replication.

On raw discrimination, no algorithm dominated. In the primary comparison, CatBoost achieved a mean precision-recall area under the curve of 0.330, narrowly ahead of ExtraTrees at 0.328; in a repeated sensitivity analysis across three independent school-grouped splits, ExtraTrees led with a mean of 0.332 while CatBoost fell to 0.296, with confidence intervals overlapping broadly. Against a random-classifier baseline of roughly 0.117 set by the base rate, these values represent nearly a three-fold improvement over chance, yet they also show that most of the variation in resilience remains unexplained. That ceiling is itself informative: it confirms the process-oriented view of resilience as context-dependent adaptation rather than a fixed, fully predictable student characteristic. CatBoost was retained for interpretation not because it was uniquely best, but because it was competitive, stable across the ten plausible-value replications, and compatible with exact TreeSHAP, a technique from explainable artificial intelligence that decomposes each prediction into contributions from individual features.

The SHAP analysis delivered the study’s most striking substantive result. The strongest predictors were not relational factors such as teacher support or school climate, as much earlier research might have predicted, but digital variables: ICT resources, digital self-efficacy, subject-related ICT use, and weekday ICT use all ranked near the top, alongside mathematics anxiety, family support, gender, and a pre-specified interaction between ICT resources and mathematics anxiety. The direction of these effects was revealing. More home ICT resources, greater confidence with technology, and more academically oriented use all pushed predictions toward resilience. But higher weekday ICT use and inquiry-oriented browsing often pushed predictions the other way. In other words, the model distinguished digital learning capacity, meaning resources that students can confidently convert into schoolwork, from sheer digital exposure, echoing the usage-gap literature showing that the quality and academic purpose of technology use, not its quantity, drives educational benefit.

Mathematics anxiety emerged as the clearest emotional marker in the profile. Low to average anxiety contributed neutrally or positively to predicted resilience, while the highest anxiety values produced the strongest negative contributions. The interaction between ICT resources and anxiety added a cross-layer twist: the benefit of digital resources was partly conditioned on emotional readiness, suggesting that anxious students may struggle to translate access into actual mathematics learning. Relational indicators did appear, but in bounded and conditional ways. Family support was present in the profile without a uniformly positive contribution, a pattern the authors read as plausibly compensatory, since support may rise when students are struggling. Teacher support and school climate indicators lacked global prominence, which the authors interpret cautiously: broad self-reported measures may vary little within a disadvantaged subgroup, and their influence may operate through classroom practices that a global feature ranking cannot capture. The negative contribution of bullying, meanwhile, is consistent with evidence that peer safety remains relevant to achievement.

Gender also surfaced as a stable, descriptive feature. Within the disadvantaged sample, weighted resilience prevalence was 9.9 percent for girls versus 14.0 percent for boys, and an ablation test showed that removing gender reduced the model’s precision-recall performance. The authors stress this does not imply any difference in inherent capability; it aligns with the national distribution, in which boys outnumbered girls among top mathematics performers, and may proxy gendered experiences of mathematics learning, expectations, or digital opportunity. Just as important is what the model cannot do. Mean recall for the resilient class was only 0.267, meaning most resilient students were missed, and the authors explicitly warn against using such predictions for individual screening, labelling, or resource allocation. The associations are descriptive and non-causal, the design is cross-sectional, and the analysis was confined to Türkiye with single imputation and without the full complex-sampling variance procedures PISA normally requires.

Still, the policy implications are pointed. In a country where the national EBA distance-learning platform and the FATİH technology infrastructure created common digital plumbing during the pandemic, the findings suggest that unequal outcomes now hinge less on connectivity than on students’ capacity to use technology for learning. Digital equity policies that stop at device provision may miss the mark; what matters is pairing access with digital self-efficacy, subject-linked use, and teacher mediation. And mathematics anxiety deserves treatment as an equity issue in its own right, particularly for students with few compensatory resources outside school. The study’s layered message is ultimately a hopeful one: resilience among disadvantaged students is not innate talent or willpower, but a configuration of conditions, and configurations, unlike traits, can be changed.

Subject of Research: Explainable machine learning analysis of academic resilience among disadvantaged students using PISA 2022 data from Türkiye

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

Article References: Erduran, M., Kuzu Demir, E. B., & Demir, K. (2026). Explainable machine learning analysis of academic resilience among disadvantaged students in Türkiye using Programme for International Student Assessment 2022 data. Discover Education, 5(1), Article 825. https://doi.org/10.1007/s44217-026-02089-2

Image Credits: AI Generated

DOI: 10.1007/s44217-026-02089-2

Keywords: academic resilience, PISA 2022, Türkiye, machine learning, SHAP, educational inequality, mathematics anxiety, digital divide, ICT use, CatBoost, socioeconomic disadvantage, explainable AI