glmnet的lasso算法对于相同的lambda参数会产生不同的结果

Question

我发现R函数glmnet::glmnet()在不同的lambda序列中，对于相同的值会产生不同的结果：

  # Generate a small dataset at random.
  set.seed(123)
  X = matrix(runif(12), ncol = 3)
  y = runif(nrow(X))
  
  
  lda = sort(c((runif(1) + 1) * 0.001, 0.001, 0.001 * (1 - runif(1))), decreasing = T)
  theModel = glmnet::glmnet(X, y, lambda = lda, intercept = T, standardize = T)
  coef(theModel)[, 2] # Show the coefficients corresponding to lambda = 0.001
  # (Intercept)          V1          V2          V3 
  # 0.8794007     1.6709897  -0.9141978  -1.5334346 
  
  
  lda = sort(c((runif(1) + 1) * 0.001, 0.001, 0.001 * (1 - runif(1))), decreasing = T)
  theModel = glmnet::glmnet(X, y, lambda = lda, intercept = T, standardize = T)
  coef(theModel)[, 2] # Show the coefficients corresponding to lambda = 0.001
  # (Intercept)          V1          V2          V3 
  # 0.8793756     1.6708981  -0.9141037  -1.5333721 

为什么会发生这种情况，如何避免它？

谢谢！

- user2961927

---

请问您能否在示例中添加适当的 set.seed(...) 以提高再现性？（这并非绝对必要，因为该现象似乎很稳健...） - Ben Bolker

Answer 1

lambda 序列将在内部进行规范化。

- user2961927