Normalization for Ridge pipelines

baraline / convst

Implementation of the Random Dilated Shapelet Transform algorithm along with interpretability tools. ReadTheDocs documentation is not up to date with the current version for now.

https://convst.readthedocs.io/en/latest/

BSD 2-Clause "Simplified" License

31 stars 7 forks source link

Normalization for Ridge pipelines #13

Closed baraline closed 2 years ago

baraline commented 2 years ago

The performance on ASCF1 is very low in a cross validation setting for RDST + Ridge, behaviour changed after deprecation of the normalize argument for Ridge classifier. TODO : Check wheter or not the "with_mean" argument on StandardScaler should be step to False. Changing it to True seems to solve the issue.

baraline commented 2 years ago

By Sklearn documentation on the normalize argument : "This parameter is ignored when fit_intercept is set to False. If True, the regressors X will be normalized before regression by subtracting the mean and dividing by the l2-norm."

Actions to close :

Change all arguments of StandardScaler to with_mean = True
Check consistency of experimental results once this is done. They are supposed to have been obtained with the use of the old sklearn interface with normalize=True.