Interpretability and feature interactions

Part 1 (2018)
1

Hi! Congratulations for your work, it is really inspiring. I have a question, in Lesson 3-4 when you work with Rossman Kaggle competitions data, you get amazing results.
But in my daily work, I usually face problems where clients do not only want to know what stores will sell more, but they also want to know why.
I have treated this problem as isolated feature importance using random forest and regressions to get the importance of each variable. My questions would be:

  • How can we interpret these complicated and complex models we build?
  • Is there a way to get the feature interaction importance? Something like stores with no nearby competitors sell a lot on weekends. I think this would be very useful for real applications, like knowing what factors combined can be the cause or a disease or things like that.

Thanks a lot once again for your incredible courses!

2

From the paper abstract:

As entity embedding defines a distance measure for categorical variables it can be used for visualizing categorical data and for data clustering.

As always the best advice is to try it out! Also make sure to check Jeremy’s paper on how feature importances from RF can be misleading.

3

Thanks a lot for your answer! I will take a look at Jeremy’s paper and will comment again!