RUA. Repositorio Institucional de la Universidad de Alicante
Publication
Otros (Other)
D-CTT: A deterministic and scalable heuristic for real-time item calibration in adaptive learning systems
Item Response Theory (IRT) offers a powerful probabilistic framework for educational assessment, but its iterative parameter estimation methods introduce computational costs and may be sensitive to convergence and estimation issues, particularly in sparse-data settings. This paper introduces D-CTT, a deterministic, non-iterative heuristic grounded in Classical Test Theory (CTT) that estimates item difficulty and subject ability through direct closed-form calculations. Unlike MMLE and JMLE approaches, whose estimates depend on iterative optimization procedures and numerical convergence, D-CTT produces exact and fully reproducible results, enabling complete algorithmic traceability of scoring decisions. This determinism also yields substantial computational advantages, making the method suitable for real-time educational applications. We evaluate D-CTT on synthetic and real-world datasets, comparing its predictive performance against Rasch model implementations based on MMLE and JMLE. An ablation study further shows that the proposed difficulty-weighting mechanism significantly improves probabilistic calibration (RMSE) over unweighted CTT, while binary accuracy remains practically equivalent. Results indicate that D-CTT achieves competitive predictive performance across a wide range of conditions and is particularly robust under high levels of data sparsity in small-sample scenarios, where iterative IRT estimators may suffer from increased variability or estimation instability. In these cases, D-CTT continues to produce valid estimates where iterative estimators stop converging, maintaining higher accuracy and lower variance. At the largest evaluated configuration, D-CTT executes over 1000 times faster than standard MMLE. These findings position D-CTT as a practical and transparent alternative for modern interactive educational systems requiring continuous parameter updates, computational efficiency, and auditable decision-making.