Rate of zeros in each feature

[sigmafelix]

We briefly discussed how to treat excessive true zeros in features even after postprocessing and imputation. The fraction of zeros per feature will give us a good reference to determine our measures.

• Before imputation, the calculation results contain 3858 columns including site_id, time, and event flag. Five lagged features will be added after imputation.
• The table below summarizes the number of features with zeros up to a certain percentile:

Percentile	50	60	70	80	90	100
Number of features	2218	2174	2100	2031	1881	1247

• To note, zero-only features were excluded before the imputation.

@kyle-messier

[sigmafelix]

The current imputation procedure includes all features unless any of these has zero variance. We could easily adjust the base function to change that.