การเปรียบเทียบวิธีพยากรณ์ในการวิเคราะห์ความถดถอยพหุคูณ โดยใช้วิธีริดจ์รีเกรสชันและวิธีที่ใช้หลักการของ โครงข่ายประสาทเทียมในกรณีที่เกิดพหุสัมพันธ์ระหว่างตัวแปรอิสระ / พัชรี คุณะสารพันธ์ = A Comparison on forecasting methods between ridge regression and artificial neural network methods in multiple regression analysis with multicollinearity
To compare the accuracy of forecasting value between multiple regression analysis of ridge regression (RR) method and artificial neural network (ANN) method when multicollinearity existing among independent variables. The criterion of comparison is the difference percentage ratio of average value of mean square error. This study used three residual distributions which are normal distribution, contaminated-normal distribution and lognormal distribution. For normal distribution, the mean of 1 and the standard deviations of 0.1, 0.3 and 0.5 are considered. For contaminated-normal distribution, the scale factors of 3 and 10, the percent of contaminations of 5 and 10 are studied. For lognormal distribution, the mean of 1, the standard deviations of 0.2264, 0.5915 and 1.0069 are tested. The sample sizes are 30, 50 and 100. When the number of independent variables is 3, the level of correlations among each pair independent variables are 0.1, 0.3, 0.5, 0.7, 0.9 and 0.99, respectively and the number of independent variables increases equal to 5, the level of correlations among independent variables (x4, x5) are 0.1, 0.3, 0.5, 0.7, 0.9 and 0.99, respectively. For each case, 400 randomly generated sets of data are used in the simulation using monte carlo technique. The result of this research can be summarized as follow. In case that residuals have nomal distribution and lognormal distribution, ranging the effect on accuracy from most to least, the accuracy of forecasting by ANN method improves as the sample size, the number of independent variables and the level of correlations among independent variables increases but decreases as the level of coefficient of variation increases. The accuracy of forecasting by RR method improves as the sample size increases but decreases as the level of correlations among independent variables, the level of coefficient of variation and the number of independent variables increases. In case the residuals have contaminated-normal distribution, ranging the effect on accuracy from most to least, the accuracy of forecasting by ANN method improves as the sample size, the number of independent variables and the level of correlations among independent variables increases but decreases as the level of coefficient of variation, the scale factors and the percent of contaminations increases. The accuracy of forecasting by RR method improves as the sample size increases but decreases as the level of correlations among independent variables, the level of coefficient of variation and the number of independent variables, the scale factors and the percent of contaminations increases. The performance of ANN method is better than that of RR method when the residuals have lognormal distribution and contaminated-normal distribution, ranging from most to least, and when the sample size, the level of coefficient of variation, the number of independent variables, the level of correlations among independent variables, the scale factors and the percent of contaminations, ranging from strongest effect to weakest effect, is larger. The performance RR method is better than that of ANN method when the residuals have normal distribution.
