[IGNITE-12685] [ML] [Umbrella] Unify Preprocessors and Pipeline approaches to collect common statistics

• Type: Improvement
• Status: Open
• Priority: Major
• Resolution: Unresolved
• Component/s: ml
• Ignite Flags: Docs Required, Release Notes Required
• Assignee: Alexey Zinoviev
• Reporter: Alexey Zinoviev
• Created: 16/Feb/20 09:43
• Updated: 29/Dec/20 08:57

Description

In the current implementation we have different behavior in Cross-Validation during running on the experimental Pipeline and chain of Preprocessors.

 

Look at the tutorial step 8 CV_Param_Grid and 8_CV_Param_Grid_and_pipeline

In the first example all preprocessors fits on the whole dataset and don't use train/test filter (due to limited API in preprocessors), and collects the stat on the whole initial dataset.

 

In the second example, we have honest re-fitting on each cross-validation fold three times with three different stats. As a result we could get a different encoding values or Max/Min values for each column and so on.

 

Should learn this question and be in consistency with the most popular approaches.