加权函数对变量筛选结果的影响

Effect of Weight Function on Variable Selection

为了明确加权函数对变量筛选结果的影响,本文以乳块消糖衣片包衣厚度为载体,用CARS方法筛选其近红外光谱特征变量,并考察加权函数对变量筛选结果的影响。结果表明,借助CARS法可以筛选出CHn在近红外区的特征性吸收变量,模型的预测性能较全谱片最小二乘(PLS)模型有一定的提高。将CARS中原加权函数(i.e.回归系数)替换为VIP后,所选变量的解释性得到提高,但模型的预测性能略有下降。即变量的解释性与模型的预测性能之间没有必然的联系。为了确保结果的可靠性,用上述方法对玉米蛋白特征变量进行筛选,可以得到类似的结论。

In this manuscript, competitive adaptive reweighted sampling （CARS） procedure was used to investigatethe effect of weight function on the selection of variables in the near-infrared spectra. A comparative study of CARS with weight function B （PLS regression coefficients） and variable importance in the projection （VIP） was conducted on two dataset taken from corn to Rukuaixiao tablets. It was shown that more pertinent variables can be extracted by CARS with built-in VIP, but the prediction performance decreased to a certain degree compared with that of synergy interval PLS （Si-PLS）.