Great question! You're right that LassoLarsCV doesn't come with a built-in random_state parameter like LassoCV does. But don't worry—you can still lock in reproducible cross-validation splits by explicitly defining a cross-validation splitter with a fixed random state, then passing it to the cv parameter of LassoLarsCV.

Here's a step-by-step breakdown to make it work:

1. Import required modules #

First, pull in the necessary tools—LassoLarsCV and a cross-validation splitter (we'll use KFold here, which matches the default behavior of LassoCV when you pass an integer to cv):

from sklearn.linear_model import LassoLarsCV
from sklearn.model_selection import KFold

2. Create a reproducible CV splitter #

Define a KFold splitter with your desired number of folds, enable shuffling (to match the random split behavior of LassoCV), and set a fixed random_state to ensure the exact same split every time you run the code:

# 3-fold cross-validation with shuffling and fixed random state
cv_splitter = KFold(n_splits=3, shuffle=True, random_state=2018)

If you don't need shuffling (and want splits based on the original data order), you can skip shuffle=True—the splits will still be reproducible as long as your input data X and y stay unchanged.

3. Fit LassoLarsCV with the custom splitter #

Pass your pre-defined splitter to the cv parameter when initializing LassoLarsCV, then fit the model as you normally would:

model_LassoLarsCV = LassoLarsCV(cv=cv_splitter).fit(X, y)

Bonus: For repeated cross-validation #

If you're using repeated cross-validation (like RepeatedKFold), the same logic applies—just set the random_state on the splitter to lock in reproducibility:

from sklearn.model_selection import RepeatedKFold

cv_splitter = RepeatedKFold(n_splits=3, n_repeats=5, random_state=2018)
model_LassoLarsCV = LassoLarsCV(cv=cv_splitter).fit(X, y)

The core idea here is that LassoLarsCV accepts any valid cross-validation generator (not just an integer) for its cv parameter. By explicitly defining a splitter with a fixed random_state, you take full control of the split process and guarantee consistent, reproducible results—just like you would with LassoCV.

内容的提问来源于stack exchange，提问作者00__00__00