可解释性
稳健性(进化)
项目反应理论
计算机科学
趋同(经济学)
蒙特卡罗方法
作弊
航程(航空)
估计
计量经济学
计算机化自适应测验
过程(计算)
差异项目功能
机器学习
数据挖掘
算法
数学优化
可靠性(半导体)
统计
项目分析
人工智能
数学
估计理论
马尔科夫蒙特卡洛
芯(光纤)
作者
Han Han,Ruihang He,Yongze Xu
标识
DOI:10.1177/01466216261489271
摘要
Item preknowledge (IP) is a severe form of cheating that fundamentally undermines the fairness and score interpretability of educational assessments. Existing detection methods for IP are predominantly constructed under the item response theory (IRT) framework, but these conventional indicators are inherently static: they rely solely on the final converged estimates of item parameters and examinee ability, while completely overlooking the rich dynamic information embedded in the iterative convergence process of ability estimation. To address this core limitation, this paper proposes a novel dynamic detection indicator-the convergence iteration count (CIC)-defined as the number of iterations required for the IRT model to converge to a stable ability estimate for each examinee. We validate the effectiveness of the CIC indicator through Monte Carlo simulated experiments across diverse scenarios, with varying proportions of compromised items and levels of misclassification uncertainty of compromised items, covering both known and unknown compromised item range conditions. Results show that the CIC indicator significantly outperforms traditional static indicators (including the TW, KL, and SR indicators) in identifying IP examinees. Notably, even in the extreme complex scenario with a 60% high proportion of compromised items and 50% high misclassification uncertainty, the area under the curve of the CIC indicator remains approaching or exceeding 0.8. The CIC indicator also exhibits strong robustness in scenarios where the range of compromised items is unknown, with detection performance surpassing most conventional non-prior-information based indicators.
科研通智能强力驱动
Strongly Powered by AbleSci AI