基于神经网络集成的IRT参数估计

Neural Networks Ensemble-Based IRT Parameter Estimation

针对统计方法难以解决小样本条件下项目反应理论(IRT)项目参数问题,提出了运用广义回归神经网络(GRNN)集成对小样本条件下项目参数进行估计的方法,运用计算机模拟的方法产生项目参数的真实值,根据双参数逻辑斯蒂模型得到被试的反应矩阵。运用经典测验理论方法得到项目难度和区分度的统计量,将其作为神经网络的输入,以模拟产生IRT的项目参数作为网络的输出,对GRNN进行训练。并且对30个神经网络加以集成,将它们在测试阶段得到输出值的平均值作为IRT参数的估计值。结果表明,神经网络集成可以得到比统计方法和单个神经网络更好的参数估计结果。

It is difficult to estimate item parameter with statistics methods based on Item Response Theory （IRT） at small sample condition. The GRNN neural networksensemble-based IRT parameter estimation method is put forward to solve this problem. The true value of item parameter are generated with computer simulation, examinees＂ response matrix is obtained based on two-parameter logistic model. The item difficulty p and discrimination r of Classical Test Theory （CTT） are used as the inputs of Generalized Regression Neural Networks （GRNN）. The simulated true value of IRT parameters are used as the outputs of GRNN. Thirty neural networks are trained and the average of their outputs in the test phase are the estimate of IRT parameters. The results shows that neural networks ensemble could get better parameter estimate than statistics method and single neural network.