Item response theory based on bayesian neural network

Abstract

Item Response Theory (IRT) is a test theory that evaluates examinees who take different tests on the same scale. However, IRT assumes random sampling examinees’ abilities from a, normal distribution. When actual examinees' abilities do not follow the normal distribution, the estimation accuracies of abilities tend to decrease significantly. To resolve this problem, Tsutsumi et.al (2021) proposed IRT model based on deep learning which estimates examinees’ abilities without assumption. However, the deep-learning-based methods tend to overfit the training data when the sample size is small. This study proposes a new IRT model, which incorporates a Bayesian neural network (BNN) into the final layer in the earlier model. BNN is a method to improve the accuracies of estimates in deep learning, by mitigating the overfitting problem. Experiments show that the proposed model improves the prediction performances of the earlier model, while it provides interoperability for both students and items.