传统的模型选择问题中,往往以所关心的性能指标的大小来评价模型.然而,许多情形下模型性能指标之间的差异在统计意义下是不显著的,是由随机误差导致的.为此,基于组块3×2交叉验证指标差异的显著性t检验给出了一种新的模型选择算法.模拟实验结果表明,在统计意义的最小化即平方损失(回归)和0-1(分类)损失最小化下,该方法常常被选择到较为简单的模型. In the traditional problem of model selection, the model is usually evaluated according to the size of the performance index concerned, however, in many cases, the difference between the model performance indexes is not statistically significant and is caused by random errors , A new model selection algorithm is given based on the significance t-test of the difference of 3 × 2 cross-validation indexes of chunks.The simulation results show that in the statistical significance minimization (squared loss (regression) and 0-1 With loss minimized, the method is often chosen to be a simpler model.