Behavioral Learning Analytics for Academic Risk Stratification in Smart Learning Platforms: A Reproducible Study Using the Public xAPI-Edu-Data Dataset

Abstract

Digital learning platforms generate rich behavioral traces that enable institutions to identify students who face academic challenges. Yet most institutional decisions continue to lack strong ties with research that can be tested and verified. The research investigates whether learning analytics createable through public xAPIEdu-Data dataset access can as-sist in determining academic risk levels for students who study through digital platforms. The dataset contains 480 student records which include 17 variables that describe their demographic and behavioral and parental and attendance characteristics. The researchers tested three classification models which included logistic regression and random forest and extra trees. The team developed a Python-based analytical pipeline which used one-hot encoding and executed five-fold crossvalidation and conducted hold-out testing. The extra trees model achieved the strongest cross-validated performance (accuracy = 0.7917, macro-F1 = 0.7959) and the best hold-out results (accuracy = 0.7986, macro-F1 = 0.8049, macro ROC-AUC = 0.9215). The feature-importance analysis revealed that student absence and resource visits and hand raises and announcement views and discussion participation and parental survey completion served as the main performance class predictors. Educational technology data provides an interpretable foundation for developing early warning systems which enable targeted academic support while conducting evidence-based student monitor-ing. The research presents three elements which include a reproducible analytical workflow and a comparative model assessment and an empirically grounded interpretation of behav-ioral indicators associated with academic performance.