我正在使用R中的glmnet包,尝试复现一个“旧”的分类器时遇到了问题。如果解释变量被排列(比如说按照相反的顺序),那么通过cv.glmnet得到的系数将不等于使用未排列矩阵计算的系数。
例如,考虑以下数据:
例如,考虑以下数据:
library(glmnet)
set.seed(1)
#Set initial parameters
n <- 100
p <- 1000
#Simulate data
x <- matrix(rnorm(n * p), nrow = n, ncol = p)
colnames(x) <- as.character(1:p)
beta <- rnorm(n = p, mean = 2, sd = 2)
beta[rbinom(p, size = 1, prob = 0.5) == 0] <- 0
y <- x %*% beta + rnorm(100, sd = 0.1)
然后对设计矩阵x和x的置换版本应用LASSO惩罚(alpha=1),并运行glmnet。
#Set parameters for cross validation with cv.glmnet
lambda <- exp(seq(-1, 1, length.out = 100))
alpha <- 1
foldid <- rep(1:10, each = 10)
#Run cross validation
fit <- cv.glmnet(x = x, y = y, family = "gaussian", alpha = alpha,
lambda = lambda,
foldid = foldid)
#Save coefficients
coef1 <- as.matrix(coef(fit, s = "lambda.min"))
#Run cross validation with rearranged design matrix
order <- ncol(x):1
fit2 <- cv.glmnet(x = x[,order], y = y, family = "gaussian", alpha = alpha,
lambda = lambda,
foldid = foldid)
#Save coefficients
coef2 <- as.matrix(coef(fit2, s = "lambda.min"))
coef2 <- coef2[rownames(coef1),]
然后比较系数、交叉验证误差和线性预测器。
> summary(coef2 - coef1)
1
Min. :-0.2738963
1st Qu.: 0.0000000
Median : 0.0000000
Mean : 0.0003739
3rd Qu.: 0.0000000
Max. : 0.3643040
> min(fit$cvm)
[1] 4584.373
> min(fit2$cvm)
[1] 4596.626
> summary(cbind(1,x) %*% coef2 - cbind(1, x) %*% coef1)
1
Min. :-0.5100
1st Qu.:-0.1613
Median : 0.0210
Mean : 0.0000
3rd Qu.: 0.1333
Max. : 0.6139
对于这三种度量标准,我们发现虽然只是变量的顺序改变了,但模型之间存在差异。有人可以解释一下吗?