Here we will use the HR churn data (https://www.kaggle.com/) to present the breakDown
package for glm models.
The data is in the breakDown package
library(breakDown)
head(HR_data, 3)
#> satisfaction_level last_evaluation number_project average_montly_hours
#> 1 0.38 0.53 2 157
#> 2 0.80 0.86 5 262
#> 3 0.11 0.88 7 272
#> time_spend_company Work_accident left promotion_last_5years sales salary
#> 1 3 0 1 0 sales low
#> 2 6 0 1 0 sales medium
#> 3 4 0 1 0 sales mediumNow let’s create a logistic regression model for churn, the
left variable.
But how to understand which factors drive predictions for a single observation?
With the breakDown package!
Explanations for the linear predictor.
library(ggplot2)
predict(model, HR_data[11,], type = "link")
#> 11
#> -0.262138
explain_1 <- broken(model, HR_data[11,])
explain_1
#> contribution
#> (Intercept) -1.601
#> satisfaction_level = 0.45 0.673
#> number_project = 2 0.568
#> salary = low 0.388
#> average_montly_hours = 135 -0.295
#> Work_accident = 0 0.221
#> time_spend_company = 3 -0.133
#> last_evaluation = 0.54 -0.129
#> promotion_last_5years = 0 0.030
#> sales = sales 0.014
#> final_prognosis -0.262
#> baseline: 0
plot(explain_1) + ggtitle("breakDown plot for linear predictors")
Explanations for the probability with intercept set as an origin.
predict(model, HR_data[11,], type = "response")
#> 11
#> 0.4348382
explain_1 <- broken(model, HR_data[11,], baseline = "intercept")
explain_1
#> contribution
#> (Intercept) 0.000
#> satisfaction_level = 0.45 0.673
#> number_project = 2 0.568
#> salary = low 0.388
#> average_montly_hours = 135 -0.295
#> Work_accident = 0 0.221
#> time_spend_company = 3 -0.133
#> last_evaluation = 0.54 -0.129
#> promotion_last_5years = 0 0.030
#> sales = sales 0.014
#> final_prognosis 1.339
#> baseline: -1.601457
plot(explain_1,
trans = function(x) exp(x)/(1+exp(x))) + ggtitle("Predicted probability of leaving the company")+ scale_y_continuous( limits = c(0,1), name = "probability", expand = c(0,0))
#> Scale for y is already present.
#> Adding another scale for y, which will replace the existing scale.