--

Good article!

As an alternative to using a specific set of functional forms for the relationships, we could use the ROC area of each feature as a "score" predicting (positively or negatively) each other feature. This is a type of nonparametric ordinal correlation. It is very effective at finding which features are related to others as a monotonic (increasing or decreasing) but otherwise-arbitrary function.

--

--

David B Rosen (PhD)

Lead Data Scientist for Automated Credit Approval (WW) at (not speaking for) IBM. linkedin.com/in/rosen1/