Explanations in natural language

knitr::opts_chunk$set(
  collapse = FALSE,
  comment = "#>",
  fig.width = 7,
  fig.height = 3.5,
  warning = FALSE,
  message = FALSE
)

Introduction

We adress the problem of insuficient interpretability of explanations for domain experts. We solve this issue by introducing describe() function, which automaticly generates natural language descriptions of explanations generated with ingredients package.

ingredients Package

The ingredients package allows for generating prediction validation and predition perturbation explanations. They allow for both global and local model explanation.

Generic function decribe() generates a natural language description for explanations generated with feature_importance(), ceteris_paribus() functions.

To show generating automatic descriptions we first load the data set and build a random forest model classifying, which of the passangers survived sinking of the titanic. Then, using DALEX package, we generate an explainer of the model. Lastly we select a random passanger, which prediction's should be explained.

library("DALEX")
library("ingredients")
library("ranger")

model_titanic_rf <- ranger(survived ~ ., data = titanic_imputed, probability = TRUE)

explain_titanic_rf <- explain(model_titanic_rf,
                            data = titanic_imputed[,-8],
                            y = titanic_imputed[,8],
                            label = "Random Forest")

passanger <- titanic_imputed[sample(nrow(titanic_imputed), 1) ,-8]
passanger

Now we are ready for generating various explantions and then describing it with describe() function.

Feature Importance

Feature importance explanation shows the importance of all the model's variables. As it is a global explanation technique, no passanger need to be specified.

importance_rf <- feature_importance(explain_titanic_rf)
plot(importance_rf)

Function describe() easily describes which variables are the most important. Argument nonsignificance_treshold as always sets the level above which variables become significant. For higher treshold, less variables will be described as significant.

describe(importance_rf)

Ceteris Paribus Profiles

Ceteris Paribus profiles shows how the model's input changes with the change of a specified variable.

perturbed_variable <- "class"
cp_rf <- ceteris_paribus(explain_titanic_rf,
                         passanger,
                         variables = perturbed_variable)
plot(cp_rf, variable_type = "categorical")

For a user with no experience, interpreting the above plot may be not straightforward. Thus we generate a natural language description in order to make it easier.

describe(cp_rf)

Natural lannguage descriptions should be flexible in order to provide the desired level of complexity and specificity. Thus various parameters can modify the description being generated.

describe(cp_rf,
         display_numbers = TRUE,
         label = "the probability that the passanger will survive")

Please note, that describe() can handle only one variable at a time, so it is recommended to specify, which variables should be described.

describe(cp_rf,
         display_numbers = TRUE,
         label = "the probability that the passanger will survive",
         variables = perturbed_variable)

Continuous variables are described as well.

perturbed_variable_continuous <- "age"
cp_rf <- ceteris_paribus(explain_titanic_rf,
                         passanger)
plot(cp_rf, variables = perturbed_variable_continuous)
describe(cp_rf, variables = perturbed_variable_continuous)

Ceteris Paribus profiles are described only for a single observation. If we want to access the influence of more than one observation, we need to describe dependence profiles.

Partial Dependence Profiles

pdp <- aggregate_profiles(cp_rf, type = "partial")
plot(pdp, variables = "fare")
describe(pdp, variables = "fare")
pdp <- aggregate_profiles(cp_rf, type = "partial", variable_type = "categorical")
plot(pdp, variables = perturbed_variable)
describe(pdp, variables = perturbed_variable)


Try the ingredients package in your browser

Any scripts or data that you put into this service are public.

ingredients documentation built on April 10, 2021, 5:06 p.m.