A Machine Learning Framework for Procrastination Detection and Academic Performance Prediction with Local Large-Language-Model Feedback

Abstract: Academic procrastination is a pervasive self-regulation failure that undermines learning outcomes, yet it is rarely detected early enough for timely intervention. Conventional academic monitoring relies on retrospective grade analysis, which identifies struggling learners only after performance has already declined. This paper presents a machine learning framework that detects procrastination behavior and predicts academic performance from passively collected study activity, then delivers individualized guidance through a locally hosted large language model. Behavioral and temporal features-session frequency, task latency, deadline proximity, and engagement regularity are extracted from learner activity logs and used to train two complementary models: an ensemble classifier that flags procrastination risk and a gradient-boosted regressor that estimates expected performance. The backend is implemented in Python, the interactive dashboard in Node.js, and contextual feedback is generated on-device by an Ollama-served language model, preserving data privacy. On a held-out evaluation, the procrastination classifier attained 92.6% accuracy and a 0.92 F1- score, while the performance regressor achieved a coefficient of determination of 0.88. A comparative study across five algorithms confirmed that gradient boosting offered the best accuracy robustness balance. The principal contributions are an interpretable behavioral feature set for procrastination modeling, a dual classification–regression pipeline that links behavior to outcomes, and a privacy-preserving local-LLM advisory layer that converts predictions into actionable, personalized recommendations for learners and educators.

Keywords: Procrastination detection; academic performance prediction; machine learning; gradient boosting; learning analytics; large language models; Ollama; educational data mining