This study investigates how students approach number-pattern problems in mathematics and whether their problem-solving strategies can be meaningfully clustered using artificial intelligence (AI) techniques. It aims to identify latent strategy profiles (e.g. guess-and-check, efficient solving, help-seeking) from digital process data to support more personalized and adaptive instruction.
This study generated synthetic student data simulating distinct strategy types and validated clustering algorithms on it. Hence, it applied the same clustering techniques to real-world student log data from ASSISTments, using features such as attempt count, hint usage and time per problem. The study evaluated cluster quality using Silhouette and Adjusted Rand Index scores and interpreted clusters through comparisons with simulated strategy types.
Five distinct student strategy profiles emerged in both simulated and real datasets. Core patterns, including efficient solvers, guessers and help-seeker, were consistently identified. Some real-data clusters aligned with the simulated types, while others revealed new, plausible strategies not captured in the initial models. This validates the clustering approach and highlights the complexity of real student behaviour.
The real-world data used had limited behavioural features (e.g. no information about diagram use or off-platform reasoning), potentially obscuring subtler strategies. Some clusters may represent blended or transitional strategies. Further studies using richer log data and triangulation (e.g. interviews, screen recordings) could improve interpretability.
Cluster awareness serves as formative assessment and enables differentiation: (i) rapid guessers – require rule articulation and self-monitoring; (ii) efficient/explicit solvers – offer extensions to deepen generalization; (iii) misconception-prone (e.g. linear projection on exponential) – prioritize early remediation; (iv) hint-forward – coach strategic help-seeking, not suppress it (Dalmaijer et al., 2022; Steinley, 2006). For learning analytics/ITS, surface interpretable indicators (error trajectories, hint sequences, pacing) and deliver profile-contingent prompts in real time (e.g. “state the rule before the next attempt”, “try a smaller step and check the rule”), with human-in-the-loop oversight and no punitive tracking (Lim, 2024). Operationally: denoise outlier dwell times, report k-selection stability, triangulate algorithms and add sequence-aware features.
For adoption, strategy profiling must centre transparency, equity and teacher empowerment. Dashboards/tutors should show which behavioural signals drive profiles, display confidence/uncertainty and prohibit punitive uses (e.g. tracking/labels without support). Institutions should monitor subgroup disparities in cluster assignment and outcomes, pair profiles with supportive – not exclusionary – interventions, and enforce privacy/consent for log data. Prioritize professional learning so educators interpret clusters responsibly and integrate them into formative practice. Such governance aligns AI-enabled analytics with instructional judgement and public trust (Pedro et al., 2019; Guan et al., 2020; Ouyang et al., 2023; Wagner, 2022).
This study combines simulation and real-data clustering to reveal latent student strategies, demonstrating how AI methods can complement educational theory. It highlights that even with coarse log features, unsupervised learning can uncover interpretable behavioural patterns. The findings contribute to advancing data-driven personalized education.
