Compute the relative importance of input variables of trained predictive models using feature shuffling
When called, the importance
function shuffles each feature n
times and computes the difference between the base score (calculated with original features X
and target variable y
) and permuted data. Intuitively that measures how the performance of a model decreases when we "remove" the feature.
- More info about the method: Permutation Feature Importance
- Permutation importance can be biased if features are highly correlated (Hooker, Mentch 2019)
// Create and train a model first
const rf = new RandomForestRegressor({
maxDepth: 20,
nEstimators: 50
})
rf.train(X, y)
// Get feature importance
const imp = importance(rf, X, y, {
kind: 'mse',
n: 10,
means: true,
verbose: false
})
console.log(impsRF)
You can also check example.js
in this repo that uses the random-forest package as a predictive model.
importance(model, X, y, options)
model
- trained model withpredict
method (predictProba
if cross-entropy used as score)X
- 2D array of featuresy
- 1D array of target variables
Options:
kind
- scoring function (mse
,mae
,rmse
,smape
,acc
,ce
(cross-entropy)n
- number of times each feature is shuffled.means
- iftrue
returns only average importanceverbose
- iftrue
throws some info into console
Feature importance is often used for variable selection. Permutation-based importance is a good method for that goal, but if you need more robust selection method check boruta.js
The importance
package is used for feature selection on StatSim Select and for data visualization on StatSim Vis