Problem Statement

Large-scale assessments are standardized evaluations conducted at regional, national, or international levels with large student populations. They support teaching and learning by enabling cross-system comparisons and providing evidence for policy decisions such as curriculum reform, teacher professional development, and equitable resource allocation (Clarke & Luna-Bazaldua, 2021; Luo, 2023; Tan et al., 2024; Zhu et al., 2025). Among these assessments, the Programme for International Student Assessment (PISA), coordinated by the OECD, is one of the most influential. Conducted every three years since 2000, PISA measures the reading, mathematics, and science literacy of 15-year-olds worldwide (Luo, 2023; OECD, 2023a; Tan et al., 2024; Zhu et al., 2025). The PISA 2022 cycle focused on mathematics literacy, defined as the ability to reason mathematically and to apply mathematics to real-life contexts (OECD, 2023a). Persistent cross-national differences in achievement suggest that mathematics literacy is shaped by educational systems, cultural orientations, and inclusion policies (Bertling & Alegre, 2019; OECD, 2023a).

Prior research has examined multiple predictors of achievement, including motivation-related factors (e.g., self-efficacy, self-concept, interest, anxiety), language background, health-related factors, socio-economic status, safety and bullying, family resources, and teacher-related factors such as guidance, fairness, feedback, and experience (Martínez-Abad & Chaparro-Caso-López, 2016; Tan et al., 2024; Wan et al., 2025; Wang et al., 2023; Zhu et al., 2025). However, multidimensional constructs such as immigration and language exposure, school culture and climate, ICT familiarity, assessment systems, teacher qualifications, and global crises remain underexplored despite their relevance for understanding equity and learning opportunities (Bertling & Alegre, 2019).

This study focuses on Germany, Switzerland, and Türkiye due to their cultural and linguistic diversity in PISA 2022. Switzerland’s multilingual structure and substantial foreign-born population create unique educational dynamics (Eurydice, 2023), while Germany and Türkiye face comparable challenges related to immigration, language integration, and inclusion. These issues gained additional importance after the COVID-19 pandemic, which disrupted learning globally and highlighted inequalities in access, preparedness, and instructional continuity (Maldonado & De Witte, 2021; UNESCO, 2022). The pandemic also increased the relevance of teacher-developed assessment practices and teachers’ digital competence for sustaining mathematics learning (Scherer et al., 2021; Wellberg, 2023).

Methodologically, PISA datasets are highly suitable for educational data-mining approaches because they include large samples, multiple countries, and a wide range of student, family, teacher, and school-level variables. Data mining, referred to as knowledge discovery in databases, involves identifying meaningful patterns, structures, and relationships within large and complex datasets (Kretowski, 2019). 

The purpose of this study is to compare the predictive performance of selected educational data-mining algorithms in classifying students’ PISA 2022 mathematics literacy performance and to identify the most influential predictors across Germany, Switzerland, and Türkiye. Also, this study aimed to examine whether PISA 2022 background variables (immigration and language exposure, school culture and climate, beliefs and attitudes, ESCS, ICT familiarity, teacher qualifications and professional development, assessment practices, and global crises) contribute to predicting students’ mathematics literacy outcomes across Switzerland, Germany, and Türkiye. Accordingly, the study addressed (1) the classification performance of C5.0 decision tree, random forest, Naive Bayes, and artificial neural network algorithms in predicting students’ PISA 2022 mathematics literacy performance and (2) the most influential predictors across the three contexts.

Method This study employed a predictive relational research design to examine factors associated with students’ mathematics literacy achievement in PISA 2022 (Creswell, 2012). Using educational data mining techniques, data from 15.524 students in Switzerland, Germany, and Türkiye were analyzed in IBM SPSS Modeler 18.0. These three countries were selected to represent high (Switzerland), middle (Germany), and low (Türkiye) performance groups. Population and Sample The sample reflected clear contextual differences across countries. Switzerland displayed the highest linguistic diversity, as students completed the assessment in German (f = 2.674; 59.8%), French (f = 1.197; 26.7%), or Italian (f = 604; 13.5%), whereas Germany (German: f = 3.964; 100%) and Türkiye (Turkish: f = 7,085; 100%) administered the assessment in a single language. The proportion of native students was lowest in Switzerland (f = 2,963; 66.2%) compared to Germany (f = 3.026; 76.3%) and Türkiye (f = 6.975; 98.4%). Home and test language differences were most common in Switzerland (f = 1.221; 27.3%), followed by Germany (f = 706; 17.8%) and Türkiye (f = 576; 8.1%). Data Collection and Variables The study used publicly available PISA 2022 mathematics literacy scores and background questionnaire scales from the official OECD database. Mathematics literacy achievement was treated as the dependent categorical variable. Independent variables included both categorical variables (LANG_DIFF, IMMIG) and continuous index variables computed by the PISA International Working Group. Data Analysis Missing data were examined prior to modeling. After excluding unanswered scales and consecutive missing responses, 15.254 observations and 28 variables were retained. The overall missing rate was 2.38%, and Little’s MCAR test indicated that data were missing completely at random (Wang et al., 2023). Random forest imputation was applied using R. Predictors were derived from weighted likelihood estimates (WLEs) and standardized for cross-country comparability. Five imputed values were generated, and their mean was used in the final dataset. The completed dataset was divided into training and test sets using a 70:30 split. Four algorithms were compared: C5.0 decision tree, random forest, Naive Bayes, and artificial neural network. Analyses were conducted separately for Türkiye, Germany, and Switzerland. Model performance was evaluated using classification metrics including accuracy, Cohen’s Kappa, AUC (Area Under the Receiver Operating Characteristic Curve), and MCC (Matthews Correlation Coefficient) (Bramer, 2020; Kretowski, 2019).

The findings showed that the C5.0 algorithm produced the highest test-data classification performance in Türkiye and Germany. In Türkiye, C5.0 correctly classified 82.78% of students, with a Kappa value of 0.568, AUC of 0.864, and MCC of 0.570. In Germany, C5.0 achieved 75.46% accuracy, with a Kappa value of 0.505, AUC of 0.833, and MCC of 0.506. In Switzerland, Naive Bayes performed best, with 74.14% accuracy, Kappa of 0.478, AUC of 0.815, and MCC of 0.477. Random forest showed relatively high training performance but lower test performance, suggesting possible overfitting. In terms of predictor importance, mathematics self-efficacy emerged as the strongest predictor in all three countries. Teacher behaviors affecting school climate, teacher-developed tests, and the difference between home language and test language were also common predictors. Country-specific patterns were also observed. In Türkiye, pandemic-related school readiness, remote instruction capacity, digital preparedness, and mathematics teacher training were prominent. In Germany, socioeconomic status, language difference, immigration status, and multicultural views were more visible. In Switzerland, mathematics self-efficacy, socioeconomic status, teacher behavior, ICT feedback, digital preparedness, and language difference formed a more distributed predictor structure.

Bertling, J. & Alegre, J. (2019). PISA 2021 context questionnaire framework (field trial version). OECD Publishing. Bramer, M. (2020). Principles of data mining. London: Springer. Clarke, M., & Luna-Bazaldua, D. (2021). Primer on large-scale assessments o educational achievement. Washington: The World Bank. Crato, N., Patrinos, H.A. (2025). PIRLS 2021 and PISA 2022 Statistics Show How Serious the Pandemic Losses Are. In: Crato, N., Patrinos, H.A. (eds) Improving National Education Systems After COVID-19. Evaluating Education: Normative Systems and Institutional Practices. Springer, Cham. https://doi.org/10.1007/978-3-031-69284-0_1 Creswell, J. W. (2012). Educational research: Planning, conducting, and evaluating quantitative and qualitative research. Boston: Pearson. Eurydice (2023). Political, social and economic background and trends. Kretowski, M. (2019). Evolutionary decision trees in large-scale data mining. Studies in Big Data. Switzerland: Springer. Luo, S. (2023). Factors affecting English reading in Macao, Hong Kong, and Singapore: combining machine learning methods and hierarchical linear regressions using pisa 2018 data (Ph. D. Dissertation). University of Macau, China Maldonado, J. E., & De Witte, K. (2021). The effect of school closures on standardised student test outcomes. British Educational Research Journal. doi:10.1002/berj.3754 Martínez-Abad, F., & Chaparro-Caso-López, A. A. (2016). Data-mining techniques i detecting factors linked to academic achievement. School Effectiveness and School Improvement, 28(1), 39-55. Organisation for Economic Co-operation and Development. (2023a). PISA 2022 results (Volume I): The state of learning and equity in education. Paris: OECD Publishing. doi:10.1787/53f23881-en. Scherer, R., Howard, S. K., Tondeur, J., & Siddiq, F. (2021). Profiling teachers’ readiness for online teaching and learning in higher education: Who’s ready? Computers in Human Behavior, 118, 1-16. Tan, L., Chen, F., & Wei, B. (2024). Examining key capitals contributing to students’ science-related career expectations and their relationship patterns: A machine learning approach. Journal of Research in Science Teaching, 61(8), 1975–2010. https://doi.org/10.1002/tea.21939 UNESCO (2022). COVID-19 School health and safety protocols: Good practices and lessons learnt to respond to Omicron. https://unesdoc.unesco.org/ark:/48223/pf0000380400.locale=en Wang, F., King, R. B., & Leung, S. O. (2023). Why do East Asian students do so well in mathematics? A machine learning study. International Journal of Science and Mathematics Education, 21(3), 691–711. https://doi.org/10.1007/s10763-022-10262-w Wellberg, S. (2023). Teacher-made tests: Why they matter and a framework for analyzing mathematics exams. Assessment in Education: Principles, Policy & Practice, 1-23. doi:10.1080/0969594X.2023.2189565 Zhu, L., You, H., Hong, M., Fang, Z. (2025). Predictive insights into U.S. students’ mathematics performance on PISA 2022 using ensemble tree-based machine learning models. International Journal of Educational Research, 130, 1-15. https://doi.org/10.1016/j.ijer.2025.102537