摘要
文章通过R软件的编程和计算,对一个帕金森病的纵向数据和一个糖尿病的横截面数据做了人工神经网络及传统参数方法的预测比较。关于第一个数据,对于训练集不同的样本量,分别运用传统的线性随机效应混合模型和神经网络做了预测,并对比标准化均方误差。发现无论是长期预测还是短期预测,线性随机效应混合模型的预测效果都显著不如神经网络。关于第二个具有多重共线性的数据,分别用岭回归、Lasso回归、适应性Lasso回归、偏最小二乘回归(PLS)、逐步回归、线性回归及神经网络方法做十折交叉验证预测对比。结果显示,神经网络在处理多重共线性数据时远远好于其他的传统参数方法,而不那么传统的PLS方法也全面优于其他几种传统方法,但远不如神经网络方法。
This paper makes use of R software programming and calculation to compare the longitudinal data of Parkinson's disease and the cross-sectional data of diabetes with the prediction of artificial neural network and traditional parameter method.About the first dataset,for the different sample sizes of the training set,the paper respectively uses the traditional linear mixed-effects(LME)model and neural network to make the prediction,and compares the standardized mean square error,discovering that the prediction effect of LME is significantly lower than that of the neural network in both the long-term prediction and the short-term prediction.About the second dataset with multicollinearity,the paper respectively uses ridge regression,Lasso regression,adaptive Lasso regression,partial least squares regression(PLS),stepwise regression,linear regression and neural network methods to make comparison of 10-fold cross validation prediction.The results show that the neural network is far better than other traditional parametric methods in processing multicollinearity data,and that the less traditional PLS method is comprehensively superior to other traditional methods,but far inferior to the neural network method.
作者
李红梅
吴喜之
王涛
Li Hongmei;Wu Xizhi;Wang Tao(School of Mathematics,Yunnan Normal University,Kunming 650092,China;School of Statistics,Renmin University of China,Beijing 100872,China)
出处
《统计与决策》
CSSCI
北大核心
2020年第9期22-25,共4页
Statistics & Decision
基金
国家自然科学基金资助项目(81360449)
清华大学横向委托项目(2014530101001536)
