Chapter 20: Communicate Model Insights

introduction

knowledge on topics varies

if basic question, explain that it is something that contributes to explain the feature and target and offer to explain later

unlocking holdout (20.1)

first check for mistakes by releasing holdout

business problem first (20.2)

done by going to data robot to unlock holdout

do the top models remain the same

how has it changed with the data process

pre-process and quality metrics (20.3)

procuring/cleaning data and addressing issues

provide algorithms used

ex. removing expired patients

provide confusion matrix and explanation of what they mean

what measure means and why its better

remove data that is unnecessary in holdout sample

extract predicted probabilities

predict then add optional features then compute predictions

areas where the model struggles (20.4)

show why the model isnt perfect

internal data (need more?)

external data (need more?)

most predictive features (20.5)

develop story and explain

not all features are created equally (20.6)

features that need to be changed and require a re-run

target leak

why it was removed

recommend business actions

features requiring further examination

possible manager follow up questions names

immutable features

good for modeling but of no value to management

mutable features

could be changed by management

will the model lose efficacy?