First of all this Feature "Importance" is different from Feature "Impact" in DataRobot
Feature Impact: measures how much the model effectiveness in prediction is affected if a feature is randomly shuffled(altered/permuted) while leaving all features unchanged
Feature Importance : measures non-linear correlation between individual predictors vs target variable
- How is the importance column calculated?
- calculated based on "Alternating Conditional Expectations" (ACE) algorithm that finds and measures the information content of the each variable in the dataset in relationship with target variable
- Does it aggregate all of the model results somehow?
- No, Because it uses a simple ACE algorithmic model.
- After we click "Start" button to build the list of models, the first thing we can notice is that "importance scores" appearing in on the green color bar even before DataRobot runs the all the models to find the championship model.
- So DataRobot can not and does not aggregate the model scores to show the "importance scores"