Can someone explain alpha and cv_alpha parameters in sklearn.linear_model.LarsCV? I am guessing that alphas refer to maximum correlation at any given step between one of the remaining explanatory variables and remaining residual. Even that might not be correct. As to the meaning of cv_alphas, that completely escapes me at the moment. Any guidance would be welcome.
So it appears that alpha is the maximum correlation (for normalized variables) between predictors and remaining residuals at a branching point where regularization path picks up a new predictor, i.e. the coefficient for a new predictor becomes nonzero.
cv_alphas have a slightly more general meaning. Each cv_alpha corresponds to some point along regularization path where maximum correlation is equal to that cv_alpha.
Just to give a little more detail, it appears that skit-learn picks a certain subrange of alphas for cross-validation. It's a subrange where it thinks the minimum cv error must reside. It then divides that subrange into nfold*nfeatures+1 points, corresponding to cv_alphas. It runs cross-validation for all those points and picks the one with the lowest cv error.
Any correction/comment would be welcome.