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

• In Event: Virtual Poster Session
  In Virtual iPresentation Only: Division D - Section 1: Measurement, Psychometrics, and Assessment Virtual Poster Session

Abstract

This paper examines the limitations of the Marginal Maximum Likelihood with Expectation-Maximization (MML-EM) algorithm for estimating 2PLM parameters in Item Response Theory (IRT), especially with limited samples and mismatched abilities and difficulties. We propose a deep learning model based on Dynamic Key-Value Memory Networks (DKVMN) and Deep-IRT, incorporating separate networks for examinees and items. Versus MML-EM, the model yields better RMSE and Pearson correlations under constrained conditions, showing promise as a cold-start estimator.

Author

• Yunbo Yang, University of Chicago