Please enable JavaScript.
Coggle requires JavaScript to display documents.
Ch 19: Interpret Model (Overall Impact of Feature w/o Considering Other…
Ch 19: Interpret Model
Feature Impacts on Target
relationships feature can have with target
1) overall impact of feature w/o considering other features
2) overall impact of feature adjusted for impact of other features
3) directional impact of feature
4) partial impact of the feature
Overall Impact of Feature w/o Considering Other Features
each feature examined individually w/ logistic regression model against target
full green bars = feature explains at least 30% of target value
green bars
pros: can focus attention on best features
cons: not fully reliable scores
features with overlapping info
should not be relied on for feature selection and model interpretation
Overall Impact of Feature Adjusted for Other Features
Feature Impact in Datarobot
randomly shuffles values of one feature within validation data segment
examines model's performance relative to model that retained all features
most important feature scored 100%
other features scaled relative to most important
fewer features is best to avoid over fitting
Directional Impact of Features on Target
whether presence of value helps model by assisting in predictions
Variable Effects in Datarobot
might want to increase running with larger sample size to increase accuracy
provides coefficients (labeled as effect)
red bars = positive cases
blue bars = negative cases
Partial Impact of Features on Target
Model X Ray in Datarobot
ranks features by their influence on target
shows 100% of whole bar
y axis = frequency
x =
The Power of Language
access to subject matter expert important during evaluation of text models
word cloud
intensity of colors indicate size of coefficients
filter stop words (words with not importance)
Hot Spots
shows most relevant combo of features and their effect on target