Educational Data Mining of Student Engagement in Science Learning: An Analytics-Based Evaluation Using Machine Learning

<ns5:p> The rapid digitalization of science education has generated extensive learning data from learning management systems (LMS), online interactions, and digital assessments. Educational Data Mining (EDM) and Learning Analytics (LA) have been widely applied to predict academic performance; however, limited studies have systematically evaluated student engagement as a multidimensional construct using integrated and multimodal analytics frameworks. Moreover, existing research often relies on single data sources, restricting comprehensive understanding. This study addresses these gaps by developing an analytics-based evaluation of student engagement in science learning that integrates explanatory and predictive approaches across multiple data sources. A mixed-method sequential explanatory design was employed. Quantitative data were collected from 126 secondary school students using LMS digital trace data, engagement surveys, and science achievement records. Structural Equation Modeling using Partial Least Squares (SEM-PLS) was applied to examine relationships among behavioral, cognitive, and interaction engagement, overall engagement, and learning outcomes. Machine learning models (Random Forest, XGBoost, Support Vector Machine, K-Nearest Neighbors, and Logistic Regression) were used to classify engagement levels and detect disengagement risk. Process mining was conducted to visualize students’ learning pathways. Qualitative data from semi-structured interviews with 12 students and 4 teachers were analyzed thematically to contextualize quantitative findings. Interaction engagement demonstrated the strongest influence on overall engagement (β = 0.33), followed by behavioral and cognitive engagement. Overall engagement significantly predicted science learning outcomes (β = 0.41; R <ns5:sup>2</ns5:sup> = 0.52), confirming its central role in academic success. Learning analytics interventions significantly enhanced engagement levels, while students in blended learning environments exhibited consistently higher engagement than those in fully online settings. Among predictive models, Random Forest achieved the highest classification accuracy (88%), with low LMS activity emerging as a key indicator of disengagement risk. Process mining revealed distinct learning trajectories, with high-engagement students exhibiting iterative and interactive learning patterns. Student engagement in science learning is multidimensional, measurable through digital data, and highly responsive to analytics-based interventions. The integration of SEM, machine learning, and process mining provides a comprehensive and robust framework for evaluating and predicting engagement. This approach offers significant implications for early warning systems, data-driven instructional design, and the optimization of blended science learning environments. </ns5:p>