Add CoxnetSurvivalAnalysisCV to search for optimal alpha via cross-validation

[plpxsk]

Curious to know if it is possible to extract the sparse features in a Cox model from L1 models.

when applying this:

pipe = make_pipeline(
    StandardScaler(),
    CoxnetSurvivalAnalysis(l1_ratio=1)
)

sfm = SelectFromModel(pipe)
sfm.fit(X_train, y_train)

n_features = sfm.transform(X).shape[1]

... I get:

ValueError: The underlying estimator Pipeline has no `coef_` or `feature_importances_` attribute.
Either pass a fitted estimator to SelectFromModel or call fit before calling transform.

source

Thanks for all the input!

[edited]

[plpxsk]

It appears that there is a coef_ method, which can be extracted as:

pipe.get_params()['coxnetsurvivalanalysis'].coef_

But this still gives error:

sfm = SelectFromModel(pipe.get_params()['coxnetsurvivalanalysis'])
sfm.fit(X_train, y_train)

n_features = sfm.transform(X_train).shape[1]

ValueError: X has a different shape than during fitting.

[sebp]

The difference is that the coef_ attribute contains the coefficients for multiple alpha values (each column corresponds to one alpha). SelectFromModel excepts a single set of coefficients, thus the exception. LassoCV is built in on top of Lasso and selects a single alpha via cross-validation. This is currently not implemented for CoxnetSurvivalAnalysis, but would be nice addition.

[plpxsk]

thanks @sebp

But then which alpha is used for the score presented from fit.score(X, y), if I keep default parameters (which sets 100 alphas)?

[Williamongh]

Thanks for your brilliant work. @sebp
And I have similar problem.
Could you please give an example for CoxnetSurvivalAnalysis ?
Thx.

BTW, have you had any idea about this problem? Thank you. @pavopax

[sebp]

The score method just calls predict without explicitly specifying an alpha value, in which case the last column of coef_ is used (corresponding to the smallest alpha in the path). If you want to compute the concordance index for multiple alpha values, you'd need to call predict and pass a value to the alpha keyword argument: model.predict(X_test, alpha=model.alphas_[1]). Of course, you can also specify alpha values not in model.alphas_ in which case coefficients will be linearly interpolated between adjacent alphas.

[CharlieCheckpt]

Hi @sebp, thank you for this great package.
I would also love to see this feature implemented; that would be very nice to have an equivalent of cv.glmnet(glmnet R package) for scikit-survival.
IMO, the main difficulty is too implement this in an efficient way, because doing cross-validation on each alpha is very costly.

[hermidalc]

@CharlieCheckpt I’m sure you’ve seen this https://gist.github.com/sebp/d580d44c4beab3379c6dfda6810b33b8

While it does GridSearchCV over each alpha in my experience since @sebp wrote CoxnetSurvivalAnalysis in C++ with Eigen it’s really fast so it doesn’t matter.

[hermidalc]

The sklearn way to determine the optimal alpha without GridSearchCV would be to design a CoxnetSurvivalAnalysisCV class very much like sklearn.linear_model.ElasticNetCV

The only disadvantage to the *CV classes in sklearn is that if you have a Pipeline composite estimator and need to do joint hyperparameter search across different steps in the pipeline then of course this must be done with GridSearchCV and you cannot use *CV estimator classes and have to instead use the non-CV versions.