Different results of CoxPHSurvivalAnalysis and CoxnetSurvivalAnalysis about scikit-survival HOT 10 CLOSED

By referencing the scikit-learn documentation (eg, see here) and experimenting with parameters, I was able to match the results of the two models:

from sksurv.linear_model import CoxPHSurvivalAnalysis, CoxnetSurvivalAnalysis
from sksurv.datasets import load_veterans_lung_cancer
from sksurv.preprocessing import OneHotEncoder

X_, y_ = load_veterans_lung_cancer()
X = OneHotEncoder().fit_transform(X_)

f = CoxPHSurvivalAnalysis(alpha=1, n_iter=100000)
g = CoxnetSurvivalAnalysis(alphas=[1/X.shape[0]], alpha_min_ratio=1, n_alphas=1, l1_ratio=1e-16, tol=1e-09, normalize=False)

print(f)
print(g)

f.fit(X, y_)
g.fit(X, y_)

print(f.coef_)
print(g.coef_[:,0])
print('Relative errors:\n', (f.coef_ - g.coef_[:,0]) / f.coef_)

Output:

CoxPHSurvivalAnalysis(alpha=1, n_iter=100000, tol=1e-09, verbose=0)
CoxnetSurvivalAnalysis(alpha_min_ratio=1, alphas=[0.0072992700729927005],
            copy_X=True, l1_ratio=1e-16, max_iter=100000, n_alphas=1,
            normalize=False, penalty_factor=None, tol=1e-09, verbose=False)
[-8.15647317e-03 -6.63327564e-01 -2.36950223e-01 -1.04657588e+00
 -3.26009359e-02 -2.71347095e-04  5.39815287e-02  2.89290932e-01]
[-8.15323389e-03 -6.63228135e-01 -2.36836564e-01 -1.04645919e+00
 -3.25993975e-02 -2.71575816e-04  5.39863418e-02  2.89279963e-01]
Relative errors:
 [ 3.97142155e-04  1.49894366e-04  4.79676516e-04  1.11497071e-04
  4.71911560e-05 -8.42911801e-04 -8.91616158e-05  3.79195919e-05]

This is reassuring.

from scikit-survival.

Now that the two models can give the same results, we can benchmark their performance and I find that CoxnetSurvivalAnalysis is about 35X faster than CoxPHSurvivalAnalysis. Given that CoxnetSurvivalAnalysis also implements a more general model (L1+L2 regularization), it makes sense to use it only for now, until one needs to dig into the details of the algorithm execution (where CoxPHSurvivalAnalysis provides better transparency).

from scikit-survival.

CoxPHSurvivalAnalysis and CoxnetSurvivalAnalysis are related but do not implement the same model. As you summarized, the difference is in penalty terms. Thus, the coefficients found by CoxPHSurvivalAnalysis and CoxnetSurvivalAnalysis will differ. In addition, CoxnetSurvivalAnalysis supports fitting a path of coefficients for varying penalization strength alpha without much computational overhead. CoxPHSurvivalAnalysis will return the result for a single a penalization strength alpha only. This is the main reason why constructors differ.

The fact that predict_survival_function is not available for CoxnetSurvivalAnalysis is an oversight which should be corrected.

from scikit-survival.

Since CoxnetSurvivalAnalysis supports L1 and L2 regularization while CoxPHSurvivalAnalysis supports only L2 regularization, by carefully matching the model parameters they should return the same results, and this is what the codes above try to do. So, why do CoxPHSurvivalAnalysis(alpha=0.5, n_iter=100000) and CoxnetSurvivalAnalysis(alphas=[0.5], alpha_min_ratio=1, n_alphas=1, l1_ratio=1e-16, tol=1e-09, normalize=False) not give the same results?

from scikit-survival.

You cannot completely eliminate the l1 penalty in CoxnetSurvivalAnalysis, thus you aren't able to exactly match the result of CoxnetSurvivalAnalysis.

from scikit-survival.

I think it makes sense to devote this issue to the different results between the two models, so I just changed the title back, and will close this thread now that it's resolved. Given that CoxnetSurvivalAnalysis seems a more desirable implementation as I concluded above, it is important to have those methods that are so far only available to CoxPHSurvivalAnalysis also available to CoxnetSurvivalAnalysis. I will open a different issue for that.

from scikit-survival.

Great work @leihuang !

from scikit-survival.

So which coef was the best recommended used for model？
the first one or the last one？
g.coef_[:,0] or g.coef_[:,-1]

from scikit-survival.

from scikit-survival.

I see. Thank you very much!

from scikit-survival.