Predicting Academic Performance from Student Behaviors and Content Engagement

Abstract

The digitalization of education has produced vast learning data, enabling AI-driven analytics to enhance teaching and learning. A key task is predicting academic performance to identify students needing extra support to pass. Much work has been done to improve the results of this task and their interpretation. However, prior research often isolates student behavior and learning materials, leading to a possible disconnect between learning content and behavior. This study bridges that gap by analyzing how students' interactions with digital content can predict academic performance. Educational log data are preprocessed to extract meaningful features that represent a combination of behavior and content, which are then refined using the Null Importance method. A LightGBM model is trained for prediction, and SHAP analysis reveals key behavioral factors linked to success. Results show that high-performing students engage more strategically and actively with critical materials. By combining behavioral and content-based analytics, this study offers a framework for early detection of learning issues and supports targeted, adaptive interventions to improve student outcomes.