iml: interpretable machine learning

iml is an R package that interprets the behaviour and explains predictions of machine learning models. It implements model-agnostic interpretability methods - meaning they can be used with any machine learning model.

Currently implemented:

• Feature importance
• Partial dependence plots
• Individual conditional expectation plots (ICE)
• Tree surrogate
• LIME: Local Interpretable Model-agnostic Explanations
• Shapley value for explaining single predictions

Read more about the methods in the Interpretable Machine Learning book

Installation

The package can be installed directly from github with devtools:

# install.packages("devtools")
devtools::install_github('christophM/iml')

Examples

First we train a randomForest to predict the Boston median housing value

library('iml')

library('randomForest')
data("Boston", package  = "MASS")
mod = randomForest(medv ~ ., data = Boston, ntree = 50)

What were the most important features? (Permutation feature importance / Model reliance)

imp = feature.imp(mod, Boston, y = Boston$medv, loss = 'mae')
plot(imp)

imp$data()

## # A tibble: 14 x 3
##    ..feature error importance
##    <fct>     <dbl>      <dbl>
##  1 age       1.28        1.30
##  2 black     1.24        1.25
##  3 chas      1.01        1.03
##  4 crim      1.65        1.67
##  5 dis       1.66        1.68
##  6 indus     1.47        1.49
##  7 lstat     4.23        4.29
##  8 medv      0.988       1.00
##  9 nox       1.64        1.66
## 10 ptratio   1.58        1.60
## 11 rad       1.11        1.12
## 12 rm        3.71        3.76
## 13 tax       1.24        1.26
## 14 zn        1.09        1.11

Let's build a single tree from the randomForest predictions! (Tree surrogate)

tree = tree.surrogate(mod, Boston[which(names(Boston) != 'medv')], maxdepth = 2)
plot(tree)

How does lstat influence the prediction on average? (Partial dependence plot)

pdp.obj = pdp(mod, Boston, feature = 13)
plot(pdp.obj)

How does lstat influence the individual predictions? (ICE)

ice.curves = ice(mod, Boston[1:100,], feature = 13)
plot(ice.curves) 

Explain a single prediction with a local linear model. (LIME)

x = Boston[1,]
lime.explain = lime(mod, Boston, x.interest = x)
lime.explain$data()

##              beta x.scaled     effect x.original feature feature.value
## rm     1.03669473    6.575  6.8162678      6.575      rm      rm=6.575
## lstat -0.05867633    4.980 -0.2922081       4.98   lstat    lstat=4.98
## medv   0.72759814   24.000 17.4623553         24    medv       medv=24

plot(lime.explain)

Explain a single prediction with game theory. (Shapley)

x = Boston[1,]
shapley.explain = shapley(mod, Boston, x.interest = x)
shapley.explain$data()

## # A tibble: 14 x 3
## # Groups:   feature [?]
##    feature      phi  phi.var
##    <fct>      <dbl>    <dbl>
##  1 age     -0.00607  0.0969 
##  2 black    0.0974   0.102  
##  3 chas    -0.00898  0.00539
##  4 crim    -0.298    1.86   
##  5 dis     -0.147    1.40   
##  6 indus    0.634    0.724  
##  7 lstat    3.12    12.7    
##  8 medv     0        0      
##  9 nox     -0.206    0.467  
## 10 ptratio  0.553    0.698  
## 11 rad     -0.156    0.0683 
## 12 rm       0.706    4.05   
## 13 tax     -0.111    0.378  
## 14 zn       0.201    0.115

plot(shapley.explain)

Python Implementation

Referring to https://github.com/datascienceinc/Skater