Simulated data

Let’s consider a following problem, the model is defined as

\[ y = x_1 * x_2 + x_2 \]

But \(x_1\) and \(x_2\) are correlated. How XAI methods work for such model?

# predict function for the model
the_model_predict <- function(m, x) {
 x$x1 * x$x2 + x$x2
}

# correlated variables 
N <- 50
set.seed(1)
x1 <- runif(N, -5, 5)
x2 <- x1 + runif(N)/100
df <- data.frame(x1, x2)

Explainer for the models

In fact this model is defined by the predict function the_model_predict. So it does not matter what is in the first argument of the explain function.

library("DALEX")
explain_the_model <- explain(1,
                      data = df,
                      predict_function = the_model_predict)

#> Preparation of a new explainer is initiated
#>   -> model label       :  numeric  ( [33m default [39m )
#>   -> data              :  50  rows  2  cols 
#>   -> target variable   :  not specified! ( [31m WARNING [39m )
#>   -> predict function  :  the_model_predict 
#>   -> predicted values  :  numerical, min =  -0.1726853 , mean =  7.70239 , max =  29.16158  
#>   -> model_info        :  package Model of class: numeric package unrecognized , ver. Unknown , task regression ( [33m default [39m ) 
#>   -> residual function :  difference between y and yhat ( [33m default [39m )
#>  [32m A new explainer has been created! [39m

Ceteris paribus

Use the ceteris_paribus() function to see Ceteris Paribus profiles. Clearly it’s not an additive model, as the effect of \(x_1\) depends on \(x_2\).

library("ingredients")
library("ggplot2")

sample_rows <- data.frame(x1 = -5:5,
                          x2 = -5:5)

cp_model <- ceteris_paribus(explain_the_model, sample_rows)
plot(cp_model) +
  show_observations(cp_model) +
  ggtitle("Ceteris Paribus profiles")

Dependence profiles

Lets try Partial Dependence profiles, Conditional Dependence profiles and Accumulated Local profiles. For the last two we can try different smoothing factors

pd_model <- partial_dependence(explain_the_model, variables = c("x1", "x2"))
pd_model$`_label_` = "PDP"

cd_model <- conditional_dependence(explain_the_model, variables = c("x1", "x2"))
cd_model$`_label_` = "CDP 0.25"

ad_model <- accumulated_dependence(explain_the_model, variables = c("x1", "x2"))
ad_model$`_label_` = "ALE 0.25"

plot(ad_model, cd_model, pd_model) +
  ggtitle("Feature effects - PDP, CDP, ALE")

cd_model_1 <- conditional_dependence(explain_the_model, variables = c("x1", "x2"), span = 0.1)
cd_model_1$`_label_` = "CDP 0.1"

cd_model_5 <- conditional_dependence(explain_the_model, variables = c("x1", "x2"), span = 0.5)
cd_model_5$`_label_` = "CDP 0.5"

ad_model_1 <- accumulated_dependence(explain_the_model, variables = c("x1", "x2"), span = 0.5)
ad_model_1$`_label_` = "ALE 0.1"

ad_model_5 <- accumulated_dependence(explain_the_model, variables = c("x1", "x2"), span = 0.5)
ad_model_5$`_label_` = "ALE 0.5"

plot(ad_model, cd_model, pd_model, cd_model_1, cd_model_5, ad_model_1, ad_model_5) +
  ggtitle("Feature effects - PDP, CDP, ALE")

Dependence profiles in groups

And now, let’s see how the grouping factor works

# add grouping variable
df$x3 <- factor(sign(df$x2))
# update the data argument
explain_the_model$data = df

# PDP in groups
pd_model_groups <- partial_dependence(explain_the_model, 
                                      variables = c("x1", "x2"), 
                                      groups = "x3")
plot(pd_model_groups) +
  ggtitle("Partial Dependence")

# ALE in groups
ad_model_groups <- accumulated_dependence(explain_the_model, 
                                      variables = c("x1", "x2"), 
                                      groups = "x3")
plot(ad_model_groups) +
  ggtitle("Accumulated Local")

# CDP in groups
cd_model_groups <- conditional_dependence(explain_the_model, 
                                      variables = c("x1", "x2"), 
                                      groups = "x3")
plot(cd_model_groups) +
  ggtitle("Conditional Dependence")

Session info

sessionInfo()

#> R version 4.0.2 (2020-06-22)
#> Platform: x86_64-apple-darwin17.0 (64-bit)
#> Running under: macOS Catalina 10.15.3
#> 
#> Matrix products: default
#> BLAS:   /Library/Frameworks/R.framework/Versions/4.0/Resources/lib/libRblas.dylib
#> LAPACK: /Library/Frameworks/R.framework/Versions/4.0/Resources/lib/libRlapack.dylib
#> 
#> locale:
#> [1] C/en_US.UTF-8/en_US.UTF-8/C/en_US.UTF-8/en_US.UTF-8
#> 
#> attached base packages:
#> [1] stats     graphics  grDevices utils     datasets  methods   base     
#> 
#> other attached packages:
#> [1] ggplot2_3.3.2       randomForest_4.6-14 ingredients_1.3.1  
#> [4] DALEX_1.3.1        
#> 
#> loaded via a namespace (and not attached):
#>  [1] knitr_1.29       magrittr_1.5     tidyselect_1.1.0 munsell_0.5.0   
#>  [5] colorspace_1.4-1 R6_2.4.1         rlang_0.4.7      dplyr_1.0.0     
#>  [9] stringr_1.4.0    tools_4.0.2      grid_4.0.2       gtable_0.3.0    
#> [13] xfun_0.15        withr_2.2.0      htmltools_0.5.0  ellipsis_0.3.1  
#> [17] yaml_2.2.1       digest_0.6.25    tibble_3.0.3     lifecycle_0.2.0 
#> [21] crayon_1.3.4     farver_2.0.3     purrr_0.3.4      vctrs_0.3.2     
#> [25] glue_1.4.1       evaluate_0.14    rmarkdown_2.3    labeling_0.3    
#> [29] stringi_1.4.6    compiler_4.0.2   pillar_1.4.6     generics_0.0.2  
#> [33] scales_1.1.1     pkgconfig_2.0.3

Simulated data, real problem

Przemyslaw Biecek

2020-07-28