Presently the FeatureImportance plot displays error bars as +/-1SD.
1) Provide an option to set this to a 95% MoE.
2) Provide a list of features that are materially positive, that is with the lower MoE > 0.
These feature importances are statistically positive with a 95% confidence based on the selected CV. If greater CV is executed the MoE will reduce. This concept makes sense when feature engineering is provided "brutally" using poly engineering on large feature sets, and we wish to trim them.
Out of scope (for this issue) is enabling integration of this component into a pipeline (a job for later).
Presently the FeatureImportance plot displays error bars as +/-1SD.
1) Provide an option to set this to a 95% MoE. 2) Provide a list of features that are materially positive, that is with the lower MoE > 0.
These feature importances are statistically positive with a 95% confidence based on the selected CV. If greater CV is executed the MoE will reduce. This concept makes sense when feature engineering is provided "brutally" using poly engineering on large feature sets, and we wish to trim them.
Out of scope (for this issue) is enabling integration of this component into a pipeline (a job for later).