In pbiecek/ceterisParibus2: Ceteris Paribus Profiles

knitr::opts_chunk$set(
  comment = "#>",
  warning = FALSE,
  message = FALSE,
  cache = TRUE
)

Here we will use the HR data to present the ceterisParibus2 package for classification models.

# devtools::install_github("ModelOriented/DALEX2")
library(DALEX2)
library(ggplot2)
library(ceterisParibus2)

head(HR)
new_observation <- HR_test[1,]
new_observation

Select neighbourhood sample, random sample, specific variables.

similar_employees <- select_neighbours(HR_test, new_observation, n = 10)
similar_employees

random_employees <- select_sample(HR_test, n = 10)
random_employees

variable_splits <- calculate_variable_split(HR, variables = c("hours", "gender"))

GLM

First, we fit a model.

library(nnet)
m_glm <- multinom(status ~ . , data = HR)

To calculete individual variable profiles (ceteris paribus profiles), i.e. series of predictions from a model calculated for observations with altered single coordinate, we use the ceterisParibus2 package.

We create an object of the ceteris_paribus_explainer class.

ivp_glm <- individual_variable_profile(m_glm,
                            data = HR_test,
                            new_observation =  new_observation)

Now we can plot individual profiles

plot(ivp_glm)

Ceteris paribus plots for neighbourhood sample, random sample, selected variables.

ivp_glm_neighbours <- individual_variable_profile(m_glm,
                            data = similar_employees,
                            new_observation =  new_observation)
plot(ivp_glm_neighbours)

ivp_glm_random <- individual_variable_profile(m_glm,
                            data = random_employees,
                            new_observation =  new_observation)
plot(ivp_glm_random)

Data frame with profiles.

profiles <- calculate_variable_profile(data = new_observation,
                                       variable_splits = variable_splits,
                                       model = m_glm)
head(profiles)

For another type of models we proceed analogously.

randomForest

library(randomForest)

m_rf <- randomForest(status ~ . , data = HR)

ivp_rf <- individual_variable_profile(m_rf,
                            data = HR_test,
                            new_observation =  new_observation)

plot(ivp_rf)

SVM

library(e1071)

m_svm <- svm(status ~ . , data = HR)

ivp_svm <- individual_variable_profile(m_svm,
                            data = HR_test,
                            new_observation =  new_observation)

plot(ivp_svm)

knn

library(caret)

m_knn <- knn3(status ~ . , data = HR, k = 5)

p_fun <- function(model, new_observation){
  predict(model, newdata = new_observation, type = "class")
}

ivp_knn <- individual_variable_profile(m_knn,
                            data = HR_test,
                            new_observation =  new_observation,
                            predict_function = p_fun)

plot(ivp_knn)

nnet

library(nnet)

m_nnet <- nnet(status ~ . , data = HR, size = 10)

p_fun <- function(model, new_observation){
  predict(model, newdata = new_observation, type = "class")
}

ivp_nnet <- individual_variable_profile(m_nnet,
                            data = HR_test,
                            new_observation =  new_observation,
                            predict_function = p_fun)

plot(ivp_nnet)

Several models at once

To produce plot with many models in one graph, use argument color = _label_.

plot(ivp_glm, ivp_rf, ivp_svm, ivp_knn, color = "_label_")

pbiecek/ceterisParibus2 documentation built on Sept. 16, 2019, 6:26 p.m.