Robust Estimation of 2PL IRT Parameters via Deep Learning under Non-Ideal Testing Conditions

This paper addresses the current limitations of the Marginal Maximum Likelihood with Expectation-Maximization (MML-EM) algorithm for parameter estimation of the 2-Parameter Logistic Model (2PLM) in Item Response Theory (IRT) under scarce item and examinee sizes, and mismatched item difficulty and examinee ability scenarios. The proposed deep learning approach, developed based on the Dynamic Key-Value Memory Network (DKVMN) and Deep-IRT, which uses separate neural networks for modeling examinees and items, outperforms the MML-EM algorithm under non-ideal data conditions of limited data sizes and mismatched examinee ability and item difficulty distributions, in terms of Root Mean Squared Error (RMSE) and Pearson correlations. The label smoothing design and modified Pearson residual loss function reinforced the model's estimation performance and interpretability. The model shows great potential for serving as a cold start method for low-resource assessment scenarios, such as dynamic testing environments, classroom testing, and promotional examinations. Future work includes further improving interpretability, model structure design for training efficiency, and conducting experiments across more diverse data conditions.