Classification example - HR data

Building a black-box model

In this example we decide to use the GBM model as a black-box - it will serve us as a surrogate.

n.trees <- 10000
library(gbm)
set.seed(111)
model_xgb1 <- gbm(left ~ ., data = data1, distribution = "bernoulli", n.trees = n.trees)

Creating an explainer

We also create an explainer object that will be used later to create new variables. For classification problems we need to specify predict_function - a function that may be used for model predictions and returns a single numerical value for each observation.

library(DALEX)
predict_function <- function(model, x) predict(model, x, type = "response", n.trees = n.trees)
explainer_xgb1 <- explain(model_xgb1, data = data2[,-7], y = data2$left, predict_function = predict_function, verbose = FALSE)

Creating a safe_extractor

Now, we create a safe_extractor object using rSAFE package and our surrogate model. Setting the argument verbose=FALSE stops progress bar from printing.

safe_extractor <- safe_extraction(explainer_xgb1, penalty = 20, verbose = FALSE)

Now, let’s print summary for the new object we have just created.

print(safe_extractor)
#> Variable 'satisfaction_level' - selected intervals:
#>  (-Inf, 0.48]
#>      (0.48, 0.9193878]
#>      (0.9193878, Inf)
#> Variable 'last_evaluation' - selected intervals:
#>  (-Inf, 0.42]
#>      (0.42, 0.57]
#>      (0.57, 0.8]
#>      (0.8, Inf)
#> Variable 'number_project' - selected intervals:
#>  (-Inf, 4]
#>      (4, Inf)
#> Variable 'average_monthly_hours' - selected intervals:
#>  (-Inf, 115.6122]
#>      (115.6122, 160]
#>      (160, 214]
#>      (214, Inf)
#> Variable 'time_spend_company' - selected intervals:
#>  (-Inf, 3]
#>      (3, Inf)
#> Variable 'work_accident' - no transformation suggested.
#> Variable 'promotion_last_5years' - no transformation suggested.
#> Variable 'salary' - created levels:
#>  high ->  high 
#>  low, medium ->  low_medium

We can see transformation propositions for all variables in our dataset.

In the plot below we can see which points have been chosen to be the breakpoints for a particular variable:

plot(safe_extractor, variable = "last_evaluation")

For factor variables we can observe in which order levels have been merged and what is the optimal clustering:

plot(safe_extractor, variable = "salary")

Transforming data

Now we can use our safe_extractor object to create new categorical features in the given dataset.

data3_trans <- safely_transform_data(safe_extractor, data3, verbose = FALSE)

We can also perform feature selection if we wish. For each original feature it keeps exactly one of their forms - original one or transformed one.

selected_variables <- safely_select_variables(safe_extractor, data3_trans, which_y = "left", verbose = FALSE)
data3_trans_sel <- data3_trans[,c("left", selected_variables)]
print(selected_variables)
#> [1] "work_accident"             "promotion_last_5years"    
#> [3] "salary"                    "satisfaction_level_new"   
#> [5] "last_evaluation_new"       "number_project_new"       
#> [7] "average_monthly_hours_new" "time_spend_company_new"

It can be observed that for some features the original form was preferred and for others the transformed one.

Here are the first few rows for our data after feature selection:

Now, we perform transformations on another data that will be used later to compare models performance.

data4_trans <- safely_transform_data(safe_extractor, data4, verbose = FALSE)
data4_trans_sel <- data4_trans[,c("left", selected_variables)]

Creating white-box models on original and transformed datasets

Let’s fit the models to data containing newly created columns. We consider a generalized linear model (glm) as a white-box model.

model_lr2 <- glm(left ~ ., data = data3_trans_sel, family = binomial())
set.seed(111)
model_xgb2 <- gbm(left ~ ., data = data3_trans_sel, distribution = "bernoulli", n.trees = n.trees)

Moreover, we create a glm model based on original data in order to check if our methodology improves results.

model_lr1 <- glm(left ~ ., data = data1, family = binomial())

Comparing models performance

Final step is the comparison of all four models we have created. For each of them we make predictions on the relevant test set, i.e. we use model_lr1 and model_xgb1 to predict the output for data2 and model_lr2 and model_xgb2 to predict the output for data4.

pred_lr1 <- round(predict(model_lr1, data2, type = "response"))
pred_xgb1 <- round(predict(model_xgb1, data2, n.trees = n.trees, type = "response"))
pred_lr2 <- round(predict(model_lr2, data4_trans_sel, type = "response"))
pred_xgb2 <- round(predict(model_xgb2, data4_trans_sel, n.trees = n.trees, type = "response"))

The performance of the models may then be evaluated based on confusion matrices with relative percentages obtained via the confusion_matrix function:

confusion_matrix <- function(y_true, y_pred) {
  cm <- data.frame(pred_0 = c(sum(y_true==0 & y_pred==0)/sum(y_true==0),
                              sum(y_true==1 & y_pred==0)/sum(y_true==1)),
                   pred_1 = c(sum(y_true==0 & y_pred==1)/sum(y_true==0),
                              sum(y_true==1 & y_pred==1)/sum(y_true==1)))
  cm <- apply(cm, MARGIN = 2, function(x) round(x, 2))
  rownames(cm) <- c("actual_0", "actual_1")
  cm
}

GBM models give the following results:

• for the original data:

• for the data with rSAFE transformations applied:

The model trained on original data has higher predictive power - feature transformations have caused the loss of valuable information which was apparently used by the first model. However, within logistic regression models the significant improvement can be observed:

• for the original data:

• for the data with rSAFE transformations applied:

The initial logistic regression model has difficulties with “true negatives” - for employees that actually quit it very often predicts the opposite. On the other hand, model_lr2 which was trained on the set of features modified by the SAFE algorithm has much better accuracy.