I prefer expectation (over typical queries) of weighted variance of the feature, divided by the total weighted variance of all features. (most likely sqrt(varia

I prefer expectation (over typical queries) of weighted variance of the feature, divided by the total weighted variance of all features. (most likely sqrt(variance) instead, but whatever power you prefer).