In ModelOriented/triplot: Explaining Correlated Features in Machine Learning Models

knitr::opts_chunk$set(
  message = FALSE, 
  warning = FALSE,
  fig.align = 'center', 
  fig.path = "man/figures/")

triplot

Introduction

The triplot package provides tools for exploration of machine learning predictive models. It contains an instance-level explainer called predict_aspects (AKA aspects_importance), that is able to explain the contribution of the whole groups of explanatory variables. Furthermore, package delivers functionality called triplot - it illustrates how the importance of aspects (group of predictors) change depending on the size of aspects.

Key functions:

• predict_triplot() and model_triplot() for instance- and data-level summary of automatic aspect importance grouping,
• predict_aspects() for calculating the feature groups importance (called aspects importance) for a selected observation,
• group_variables() for grouping of correlated numeric features into aspects.

The triplot package is a part of DrWhy.AI universe. More information about analysis of machine learning models can be found in the Explanatory Model Analysis. Explore, Explain and Examine Predictive Models e-book.

Installation

# from CRAN:
install.packages("triplot")

# from GitHub (development version):
# install.packages("devtools")
devtools::install_github("ModelOriented/triplot")

Overview

triplot shows, in one place:

• the importance of every single feature,
• hierarchical aspects importance,
• order of grouping features into aspects.

We can use it to investigate the instance level importance of features (using predict_aspects() function) or to illustrate the model level importance of features (using model_parts() function from DALEX package). triplot can be only used on numerical features. More information about this functionality can be found in triplot overview.

Basic triplot for a model

To showcase triplot, we will choose apartments dataset from DALEX, use it's numeric features to build a model, create DALEX explainer, use model_triplot() to calculate the triplot object and then plot it with the generic plot() function.

Import apartments and train a linear model

library("DALEX")
apartments_num <- apartments[,unlist(lapply(apartments, is.numeric))]

model_apartments <- lm(m2.price ~ ., data = apartments_num)

Create an explainer

explain_apartments <- DALEX::explain(model = model_apartments, 
                              data = apartments_num[, -1],
                              y = apartments_num$m2.price,
                              verbose = FALSE)

Create a triplot object

set.seed(123)
library("triplot")

tri_apartments <- model_triplot(explain_apartments)

plot(tri_apartments) + 
  patchwork::plot_annotation(title = "Global triplot for four variables in the linear model")

At the model level, surface and floor have the biggest contributions. But we also know that Number of rooms and surface are strongly correlated and together have strong influence on the model prediction.Construction year has small influence on the prediction, is not correlated with number of rooms nor surface variables. Adding construction year to them, only slightly increases the importance of this group.

Basic triplot for an observation

Afterwards, we are building triplot for single instance and it's prediction.

(new_apartment <- apartments_num[6, -1])

tri_apartments <- predict_triplot(explain_apartments, 
                                  new_observation = new_apartment)

plot(tri_apartments) + 
  patchwork::plot_annotation(title = "Local triplot for four variables in the linear model")

We can observe that for the given apartment surface has also significant, positive influence on the prediction. Adding number of rooms, increases its contribution. However, adding construction year to those two features, decreases the group importance.

We can notice that floor has the small influence on the prediction of this observation, unlike in the model-level analysis.

Aspect importance for single instance

For this example we use titanic dataset with a logistic regression model that predicts passenger survival. Features are combined into thematic aspects.

Importing dataset and building a logistic regression model

set.seed(123)

model_titanic_glm <- glm(survived ~ ., titanic_imputed, family = "binomial")

Manual selection of aspects

aspects_titanic <-
  list(
    wealth = c("class", "fare"),
    family = c("sibsp", "parch"),
    personal = c("age", "gender"),
    embarked = "embarked"
  )

Select an instance

We are interested in explaining the model prediction for the johny_d example.

(johny_d <- titanic_imputed[2,])

predict(model_titanic_glm, johny_d, type = "response")

It turns out that the model prediction for this passenger's survival is very low. Let's see which aspects have the biggest influence on it.

We start with DALEX explainer.

explain_titanic <- DALEX::explain(model_titanic_glm, 
                           data = titanic_imputed,
                           y = titanic_imputed$survived,
                           label = "Logistic Regression",
                           verbose = FALSE)

And use it to call triplot::predict_aspects() function. Afterwards, we print and plot function results

library("triplot")

ai_titanic <- predict_aspects(x = explain_titanic, 
                              new_observation = johny_d[,-8],
                              variable_groups = aspects_titanic)

print(ai_titanic, show_features = TRUE)

plot(ai_titanic)

We can observe that wealth (class, fare) variables have the biggest contribution to the prediction. This contribution is of a negative type. Personal (age, gender) and Family (sibsp, parch) variables have positive influence on the prediction, but it is much smaller. Embarked feature has very small, negative contribution to the prediction.

Learn more

• triplot package overview
• usecase with FIFA 20 data set
• description of predict aspects method

Acknowledgments

Work on this package was financially supported by the NCBR Grant POIR.01.01.01-00-0328/17.

ModelOriented/triplot documentation built on March 10, 2021, 6:26 p.m.