Document Type : Applied Article
Authors
Department of Management, Faculty of Social and Economic Sciences, Alzahra University, Tehran, Iran.
Abstract
Risk screening in large-scale educational assessment can support the early identification of students who have not reached minimum performance benchmarks. This study develops a weighted and interpretable machine-learning framework for classifying fourth-grade Iranian students at risk of low science achievement using TIMSS 2023 Iran data. To avoid a sample-dependent risk definition, students scoring below 400 were classified as at risk, corresponding to those who did not reach the TIMSS Low International Benchmark. Thirteen individual, socioeconomic, and home learning resource predictors were used. To reduce school-level information leakage, the data were partitioned by school into training, validation, and test sets. A feedforward neural network was trained using cross-entropy loss with error weights combining TIMSS sampling weights and class-balancing weights. Model evaluation used sampling-weighted metrics, and the classification threshold was selected on the validation set by maximizing the weighted F1-score. Ridge logistic regression, decision tree, random forest, and boosted-tree models were used as benchmarks. On the test set, the neural network achieved a recall of 0.833, F1-score of 0.682, and weighted AUC of 0.739, while ridge logistic regression achieved slightly higher F1-score and weighted AUC. Grouped permutation importance highlighted parental education, household affluence, number of books at home, language background, and home learning resources as influential predictors. The findings suggest that benchmark-based, interpretable risk screening can support preliminary educational decision-making when used as a decision-support tool rather than a definitive labeling system.
Keywords
Main Subjects
)