From a pedagogical point of view, I find it weird to do an example on automated feature selection on a dataset which was prepared with manual feature selection.

However, since the example does not really rely on the choice of the Lasso penalty, maybe we could use just a faster base linear model, for instance RidgeCV(alphas=np.logspace(-6, 6, num=5)) on the original feature set.

However, since the example does not really rely on the choice of the Lasso penalty, maybe we could use just a faster base linear model, for instance RidgeCV(alphas=np.logspace(-6, 6, num=5)) on the original feature set.

I agree with @ogrisel. In the original example, the penalty chosen will let out only a single feature. So using RidgeCV will be much cheaper and keep the description more or less the same. I think that what is really expensive here is indeed the SequentialFeatureSelector. We could reduce to cv=3 and check if this is enough (maybe with an additional note to mention that in practice one should consider to increase the number of folds)

Using RidgeCV the results are much faster on the original feature set:

Features selected by SelectFromModel: ['s1' 's5']
Done in 0.001s
Features selected by forward sequential selection: ['bmi' 's5']
Done in 0.116s
Features selected by backward sequential selection: ['bmi' 's5']
Done in 0.305s
real 6.60
user 2.20
sys 0.88

Graph generated:

Should I move forward and update the file?

Using RidgeCV the results are much faster on the original feature set:

Yes this is expected. Can you make the change?

The test failure is unrelated.

Thanks @ArthDh Merging since the failure in the CI is unrelated with your changes.