iml
Interpretable Machine Learning.
Create Predictor
library(iml)
# Create predictor object
predictor <- Predictor$new(
model = model,
data = train_data,
y = train_labels
)
Feature Importance
# Permutation importance
imp <- FeatureImp$new(predictor, loss = "mse")
plot(imp)
# Results
imp$results
Partial Dependence
# Single feature
pdp <- FeatureEffect$new(predictor, feature = "age")
plot(pdp)
# Method options
pdp <- FeatureEffect$new(predictor, feature = "age",
method = "pdp") # Partial dependence
pdp <- FeatureEffect$new(predictor, feature = "age",
method = "ale") # Accumulated local effects
pdp <- FeatureEffect$new(predictor, feature = "age",
method = "pdp+ice") # PDP + ICE
Feature Interactions
# Two-way interaction
interact <- Interaction$new(predictor)
plot(interact)
# Specific feature
interact <- Interaction$new(predictor, feature = "age")
SHAP Values
# Shapley values for single prediction
shap <- Shapley$new(predictor, x.interest = new_data[1, ])
plot(shap)
# Results
shap$results
LIME
# Local interpretable model
lime <- LocalModel$new(predictor, x.interest = new_data[1, ])
plot(lime)
# Results
lime$results
Surrogate Model
# Global surrogate
tree <- TreeSurrogate$new(predictor, maxdepth = 3)
plot(tree)
# Predict with surrogate
tree$predict(new_data)
Counterfactuals
# What-if counterfactuals
library(counterfactuals)
cf <- Counterfactuals$new(predictor, x.interest = new_data[1, ])
cf$find_counterfactuals(desired_outcome = 1)
Multiple Features
# 2D PDP
pdp2d <- FeatureEffect$new(predictor,
feature = c("age", "income"),
method = "pdp")
plot(pdp2d)