Using In-Class Exercise Data for Early Support of Struggling Students

Abstract

This classroom experience report investigates how in-class programming exercises can function as early-warning systems for struggling students. We studied two graduate Computer Science courses (n = 57) where students’ code was automatically captured every 20 seconds during short programming activities. These fine-grained behavioral snapshots created a continuous record of real-time engagement. Clustering analysis revealed three consistent profiles: Early Birds (44.4%) began quickly and achieved strong outcomes, Active Strugglers (33.3%) engaged steadily but solved fewer problems, and Delayed/Disengaged students (22.2%) delayed by 10+ minutes and performed poorly. These profiles emerged within the first six weeks of the semester and strongly predicted exam performance (exercise access delay correlated with midterm scores, r = −0.68). To explore timely intervention, we developed personalized practice materials for struggling students based on their observed patterns. Large language models (LLMs) were used to analyze code snapshots and generate tailored exercises. These materials were then reviewed for technical accuracy and pedagogical alignment before delivery. Among the 11 students receiving this support, 81.8% achieved high-performer status (≥ 85% on both exams), substantially exceeding baseline expectations. This study contributes to research on early warning systems, procrastination in computing education, and fine-grained behavioral analytics, while also demonstrating how LLMs can be integrated into classroom practice. Our findings suggest that routine programming activities can serve a dual purpose: supporting active learning while simultaneously providing a scalable, low-cost framework for early identification and intervention.