Which feature importance methods does the data robot platform use? How can I choose between different methods, if there's one I want to use? thanks for helping me
Hi @data-dino, I'd like to ask a clarifying question. What are you planning on doing once you look at feature importance? What decision will you make based on what you discover?
If I understand that I can help recommend what might be best for you.
Hi! Datarobot provides a variety of insights for features that may be addressed as feature importance:
Feature importance: ACE score built during EDA (exploratory data analysis) as mentioned above
Feature associations: Feature clusters, feature association matrix, feature association pair plots.
Feature Impact: built using a trained model, Permutation based, SHAP-based
Feature Effects: provide model detail insights on a per-feature basis
Insights: Tree-based, Coefficients-based, Word clouds, etc.
Overall, DataRobot recommends using either permutation-based or SHAP-based Feature Impact as they show results for original features in predictive models. Hope this answer will help. If you need further explanations - feel free to ask.
Hi @data-dino , the feature importance method we use is called "Alternating Conditional Expectation (ACE)". It is a univariate measure of correlation between the feature and the target. ACE scores detect non-linear relationships, but as they are univariate, they do not detect interaction effects. If you want to go into more details you can have a look to this paper: https://www.jds-online.com/files/JDS-156.pdf
You can't choose a different method. Out of curiosity, which method you have in mind ?
日本語
