Wheel-spinning refers to a pattern of unproductive persistence in which students repeatedly practice a skill without achieving mastery, despite receiving feedback and multiple opportunities for correction. This phenomenon poses a significant challenge for intelligent tutoring systems (ITSs) and adaptive learning environments, as delayed identification of struggling learners can lead to frustration, disengagement, and attrition. Although prior studies have proposed various wheel-spinning detectors, many existing approaches rely on complex models, extensive response histories, or system-specific features, limiting their scalability and early-detection potential.
This study develops an early and simplified wheel-spinning detection model that relies exclusively on student response-sequence features. Using data from the Cognitive Model Discovery Experiment (Spring 2010) in the Carnegie Mellon University DataShop repository, the analysis included 123 students, 49 geometry skills, and over 45,000 response records. Wheel-spinning was operationalized as failure to achieve mastery—defined as three consecutive correct first attempts—within ten practice opportunities. Two models were compared: a baseline logistic regression model and an enhanced ensemble model using gradient boosted decision trees (GBDT).
Results indicate that both models achieved strong predictive performance at early practice stages. The logistic regression model demonstrated high accuracy and precision but moderate recall, identifying confirmed wheel-spinning cases while missing a subset of struggling students. In contrast, the GBDT model substantially improved recall while maintaining high accuracy, achieving over 70% accuracy and approximately 75% recall by the fifth practice opportunity. These findings suggest that ensemble learning methods are particularly effective for early identification of unproductive persistence.
Overall, the study demonstrates that lightweight, response-based models can detect wheel-spinning early and reliably without reliance on tutor-specific parameters. The proposed approach offers a scalable and interpretable solution for adaptive learning systems, enabling timely instructional interventions and more responsive learner support.
