Permutation Variable Importance

Introduction

Permutation variable importance is obtained by measuring the distance between prediction errors before and after a feature is permuted; only one feature at a time is permuted.

Implementation

The model is scored on a dataset D, this yields some metric value orig_metric for metric M.

Permutation variable importance of a variable V is calculated by the following process:

  1. Variable V is randomly shuffled using Fisher-Yates algorithm.

  2. The model is scored on the dataset D with the variable V replaced by the result from step 1. this yields some metric value perm_metric for the same metric M.

  3. Permutation variable importance of the variable V is then calculated as abs(perm_metric - orig_metric).

Metric M can be set by metric argument. If set to AUTO, AUC is used for binary classification, logloss is used for multinomial classification, and RMSE is used for regression.

Parameters

  • model: A trained model for which it will be used to score the dataset.

  • frame: The dataset to use, both train and test frame are can be reasonable choices but the interpretation differs (see Should I Compute Importance on Training or Test Data? from the Interpretable Machine Learning by Christoph Molnar.).

  • metric: The metric to be used to calculate the error measure. One of AUTO, AUC, MAE, MSE, RMSE, logloss, mean_per_class_error, PR_AUC. Defaults to AUTO.

  • n_samples: The number of samples to be evaluated. Use -1 to use the whole dataset. Defaults to 10 000.

  • n_repeats: The number of repeated evaluations. Defaults to 1.

  • features: The features to include in the permutation importance. Use None to include all.

  • seed: The seed for the random generator. Use -1 to pick a random seed. Defaults to -1.

Output

When n_repeats == 1, the result is similar to the one from h2o.varimp(), i.e., it contains the following columns “Relative Importance”, “Scaled Importance”, and “Percentage”.

When n_repeats > 1, the individual columns correspond to the permutation variable importance values from individual runs which corresponds to the “Relative Importance” and also to the distance between the original prediction error and prediction error using a frame with a given feature permuted.

Examples

library(h2o)

# start h2o
h2o.init()

# load data
prostate_train <- h2o.importFile("http://s3.amazonaws.com/h2o-public-test-data/smalldata/prostate/prostate.csv")

# train a model
gbm <- h2o.gbm(y = "CAPSULE", training_frame = prostate_train)

# calculate importance
permutation_varimp <- h2o.permutation_importance(gbm, prostate_train, metric = "MSE")

# plot permutation importance (bar plot)
h2o.permutation_importance_plot(gbm, prostate_train)

# plot permutation importance (box plot)
h2o.permutation_importance_plot(gbm, prostate_train, n_repeats=15)
import h2o
from h2o.estimators import *

# start h2o
h2o.init()

# load data
prostate_train = h2o.import_file("http://s3.amazonaws.com/h2o-public-test-data/smalldata/prostate/prostate.csv")
prostate_train["CAPSULE"] = prostate_train["CAPSULE"].asfactor()

# train model
gbm = H2OGradientBoostingEstimator()
gbm.train(y="CAPSULE", training_frame=prostate_train)

# calculate importance
permutation_varimp = gbm.permutation_importance(prostate_train, use_pandas=True)

# plot permutation importance (bar plot)
gbm.permutation_importance_plot(prostate_train)

# plot permutation importance (box plot)
gbm.permutation_importance_plot(prostate_train, n_repeats=15)