GaussianProcess score function #29
Conversation
- a few cosmits in the examples - implementation of real built-in score function for the GaussianProcess which evaluates a leave-one-out estimate of the determination coefficient (as usually done w/ this kind of model).
|
@dubourg what is the status of this pull request? |
Conflicts: examples/gaussian_process/plot_gp_diabetes_dataset.py examples/gaussian_process/plot_gp_probabilistic_classification_after_regression.py
|
merged ! thanks |
|
Thanks for the mege. But I'd like to review the implementation of the cross_val_score function to address the problem discussed last night: one cannot perform a leave-one-out estimate on the default Regressor's score function as inhereted from base.RegressorMixin. Indeed, the r2_score(y_true, y_pred) would return a -Inf array as it computes the variance on a single sample, thus y_true -y_true.mean() = 0 and finally R2 = -Inf... I think that a cross_val_predict function could do the trick for all regressors in the scikits, and it would allow to make adequation plots: y_pred_on_folds vs. y_true in addition to the cross_val score estimates. |
FIX: some of Gael comments
ENH : NumpyDocString subclass collections.Mapping
Hi list,
After a discussion with Alex (Gramfort) who reviewed the gaussian_process module, we decided to re-implement the score function so that it natively performs a leave-one-out estimate of the determination coefficient (as usually done with this kind of model) without exposing the internals to the user.
(See agramfort@f068df4)
There is an example of usage of this new score function in the plot_gp_diabetes_dataset.py example.
Cheers,
Vincent