Explain

The explore package offers a simplified way to use machine learning to understand and explain patterns in the data.

library(dplyr)
library(explore)

data <- create_data_buy(obs = 1000)
glimpse(data)
#> Rows: 1,000
#> Columns: 13
#> $ period          <int> 202012, 202012, 202012, 202012, 202012, 202012, 202012…
#> $ buy             <int> 1, 0, 1, 0, 0, 1, 0, 0, 0, 1, 0, 1, 0, 0, 1, 0, 1, 1, …
#> $ age             <int> 39, 57, 55, 66, 71, 44, 64, 51, 70, 44, 58, 47, 68, 71…
#> $ city_ind        <int> 1, 0, 0, 0, 0, 1, 0, 1, 1, 1, 0, 0, 1, 1, 0, 1, 1, 1, …
#> $ female_ind      <int> 0, 0, 1, 0, 1, 0, 0, 0, 0, 1, 1, 0, 0, 0, 1, 0, 1, 0, …
#> $ fixedvoice_ind  <int> 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, …
#> $ fixeddata_ind   <int> 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, …
#> $ fixedtv_ind     <int> 1, 0, 1, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 0, 1, 1, 1, 0, …
#> $ mobilevoice_ind <int> 0, 1, 1, 0, 0, 1, 0, 1, 1, 1, 0, 1, 0, 1, 1, 1, 0, 0, …
#> $ mobiledata_prd  <chr> "NO", "NO", "MOBILE STICK", "NO", "BUSINESS", "BUSINES…
#> $ bbi_speed_ind   <int> 1, 1, 0, 1, 0, 1, 0, 1, 0, 1, 1, 1, 0, 0, 1, 0, 0, 1, …
#> $ bbi_usg_gb      <int> 77, 49, 53, 44, 55, 93, 50, 64, 63, 87, 45, 45, 70, 79…
#> $ hh_single       <int> 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 0, 0, 1, 0, 0, 0, 0, 1, …

Explain / Model

Decision Tree

data %>% explain_tree(target = buy)

data %>% explain_tree(target = mobiledata_prd)

data %>% explain_tree(target = age)

Random Forest

data %>% explain_forest(target = buy, ntree = 100)

To get the model itself as output you can use the parameter out = "model or out = all to get all (feature importance as plot and table, trained model). To use the model for a prediction, you can use predict_target()

XGBoost

As XGBoost only accepts numeric variables, we use drop_var_not_numeric() to drop mobile_data_prd as it is not a numeric variable. An alternative would be to convert the non numeric variables into numeric.

data %>%
  drop_var_not_numeric() |> 
  explain_xgboost(target = buy)

Use parameter out = "all" to get more details about the training

train <- data %>%
  drop_var_not_numeric() |> 
  explain_xgboost(target = buy, out = "all")

train$importance
#>           variable        gain       cover  frequency  importance
#>             <char>       <num>       <num>      <num>       <num>
#> 1:             age 0.480076565 0.346517287 0.30325098 0.480076565
#> 2:      bbi_usg_gb 0.263936257 0.306661425 0.31781534 0.263936257
#> 3:      female_ind 0.133861444 0.138796555 0.12197659 0.133861444
#> 4:     fixedtv_ind 0.082409865 0.106854333 0.10221066 0.082409865
#> 5:        city_ind 0.020369882 0.051517918 0.05539662 0.020369882
#> 6:   bbi_speed_ind 0.010279095 0.020894635 0.03276983 0.010279095
#> 7:       hh_single 0.005200283 0.014226769 0.03250975 0.005200283
#> 8: mobilevoice_ind 0.002025957 0.006488770 0.01742523 0.002025957
#> 9:  fixedvoice_ind 0.001840650 0.008042308 0.01664499 0.001840650

train$tune_plot

train$tune_data
#>    model_nr   eta max_depth    runtime  iter train_auc_mean test_auc_mean
#>       <int> <num>     <num> <difftime> <int>          <num>         <num>
#> 1:        1  0.30         3     0 mins    29      0.9670353     0.9229369
#> 2:        2  0.10         3     0 mins    54      0.9585660     0.9259053
#> 3:        3  0.01         3     0 mins   591      0.9602903     0.9265414
#> 4:        4  0.30         5     0 mins    12      0.9738282     0.9218968
#> 5:        5  0.10         5     0 mins    21      0.9674193     0.9250532
#> 6:        6  0.01         5     0 mins    25      0.9535213     0.9242371

To use the model for a prediction, you can use predict_target()

Logistic Regression

data %>% explain_logreg(target = buy)
#> # A tibble: 6 × 5
#>   term          estimate std.error statistic  p.value
#>   <chr>            <dbl>     <dbl>     <dbl>    <dbl>
#> 1 (Intercept)  5.87      0.544        10.8   3.88e-27
#> 2 age         -0.146     0.0106      -13.8   3.49e-43
#> 3 city_ind     0.711     0.183         3.89  1.02e- 4
#> 4 female_ind   1.75      0.186         9.38  6.91e-21
#> 5 fixedtv_ind  1.51      0.190         7.93  2.14e-15
#> 6 bbi_usg_gb  -0.0000724 0.0000904    -0.801 4.23e- 1

Balance Target

If you have a data set with a very unbalanced target (in this case only 5% of all observations have buy == 1) it may be difficult to create a decision tree.

data <- create_data_buy(obs = 2000, target1_prob = 0.05)
data %>% describe(buy)
#> variable = buy
#> type     = integer
#> na       = 0 of 2 000 (0%)
#> unique   = 2
#>        0 = 1 899 (95%)
#>        1 = 101 (5.1%)

It may help to balance the target before growing the decision tree (or use weighs as alternative). In this example we down sample the data so buy has 10% of target == 1.

data %>%
  balance_target(target = buy, min_prop = 0.10) %>%
  explain_tree(target = buy)