Comparative Research of BP Neural Network Estimates IRT Parameters
Lingli Chen · Zhongguo kaoshi · 2013
Objective : Compared with classical test theory, item response theory has more advantages, However, item response theory models are complex , the parameter estimates are often required a large sample ; Artificial neural network may provide methods for small sample to estimate item response theory parameters , The purpose of the article is to find a more accurate parameter estimation by Monte Carlo simulation of neural networks . Method : two parameters item response theory model as an example , MAB and RMSE as compared indicators , Comparison the differences of values by simulation data between percentage , point-biserial correlation coefficient, the average score in the Classical Test Theory and the converted value (IRT parameter estimation of initial value) as the neural network input values in neural network training network , compare two indicators RMSE and MAB index under different condition . Result : there is a difference between item percentage and b j =z j/r bj in estimating the item parameters b ; there is a difference between point-biserial correlation coefficient and a j =r bj/1-r 2 bj in estimating item parameters a ; there is a difference between average score and ln[ x/(m-x)] in estimating ability parameters theta. Conclusion : For the two-parameter item response model , error in item percentage is smaller than b j =z j/r bj in estimating item parameter b ; error in point-biserial correlation coefficient is larger than a j =r bj /1-r 2 bj in estimating item parameter a ; error in average score is larger than ln[ x/(m-x)] in estimating ability parameter theta.