Tracing learner responsiveness through temporal interaction patterns

Abstract

Educational AI systems increasingly seek to personalize support by tracking how learners interact with feedback and instructional events. However, many current systems emphasize activity quantity over learners’ responsiveness to pedagogical timing. This study models behavioural responsiveness, operationalized as learner clicks occurring within 24 h of machine-generated events, using log data from 10,452 learners in a large-scale online platform. We introduce novel temporal features, including burstiness, entropy, and the Human–Machine Collaboration Index (HMCI), a metric of behavioural synchrony between learner activity and system events. Four machine learning models (logistic regression, random forest, XGBoost, and neural networks) classified course withdrawal and behavioural responsiveness with high internal accuracy and interpretability (e.g., SHAP analysis). Importantly, these models are not intended for real-time prediction but rather to evaluate the coherence of a behavioural proxy. Findings suggest that modelling temporal alignment, rather than raw engagement, offers a promising direction for designing responsive, interpretable educational technologies.