Interpret and Communicate
Interpret Model
Feature Impacts on Target
Overall impact on features w/o consideration for other features
Overall impact adjusted for other features
directional impact of a feature
partial impact of the feature
Standalone effect on target
Green bar
Log loss
Most likely to yield predictive value
many will contain same information
Each feature adjusted for the other features in the model
scaled from 1-100
Usain Bolt's individual limbs
fewer features help prevent overfitting
Insights
Variable effects
Regression analysis
16% not well enough for 32 or 64%
Processing time vs accuracy
Re-admits vs non-readmits
Directional impact variation
Model xRay
ranks by influence
discharge disposition iD
orange dot represents actual socre
Blue dot labeled predicted
data robot is crafted to prevent overfiitting
partial dependence
numeric features
The Power of Language
Bring the SME Back into it
easier to train SME in ML than vice versa
filter stop words
predictive of
Hotspots
Rulefit classifier
Run with new sample size
4 Most relevant
mean relative target
average number of readmits
How accurate the hot spot is
Reason code
data validation set
left and right validation sets