Added models trained on raw dataset (without binning) for both Ranger…

… and XGBoost models. Automated code in "rand_forest.R", it's now training two models by default.

dataset_prep.R

@@ -17,6 +17,10 @@ data1 <- data_raw %>%
          HasCrCard = factor(HasCrCard) %>% `levels<-`(c("No", "Yes"))) %>% 
   dplyr::select(-RowNumber, -CustomerId, -Surname)
 
+data1 %>% filter_vars_by_iv(significance_thres = 0.02) %>% 
+  initial_split(prop = 0.75) %>%
+  saveRDS("data/split_raw.RDS")
+
 data2 <- data1 %>% 
   factorize(bin_methods = "tree") %>% 
   as_tibble() %>% 

gbm.R

@@ -11,15 +11,17 @@ rm(list = ls())
 # source("dataset_prep.R")
 
 dataset_split <- readRDS("data/split.RDS")
-dataset_split$data <-  dataset_split$data %>% 
-  mutate_if(~ length(levels(.x)) > 3, as.integer) %>% 
+dataset_split$data <-  dataset_split$data %>%
+  mutate_if(~ length(levels(.x)) > 3, as.integer) %>%
   mutate_at(vars(Balance), as.integer)
 
+# dataset_split <- readRDS("data/split_raw.RDS")
+
 df_train <- dataset_split %>% training()
 df_test <- dataset_split %>% testing()
 
 gbm_model_1 <- boost_tree(mode = "classification",
-                          mtry = 2,
+                          mtry = 3,
                           trees = 500,
                           min_n = 5,
                           # tree_depth = 5,

rand_forest.R

@@ -10,32 +10,52 @@ rm(list = ls())
 
 # source("dataset_prep.R")
 
-dataset_split <- readRDS("data/split.RDS")
+dataset_split1 <- readRDS("data/split.RDS")
+dataset_split2 <- readRDS("data/split_raw.RDS")
 
-df_train <- dataset_split %>% training()
-df_test <- dataset_split %>% testing()
+df_train1 <- dataset_split1 %>% training()
+df_test1 <- dataset_split1 %>% testing()
+df_train2 <- dataset_split2 %>% training()
+df_test2 <- dataset_split2 %>% testing()
 
-ranger_model_1 <- rand_forest("classification", 2, 1000, 5) %>% 
-  set_engine("ranger", num.threads = 8, replace = F, sample.fraction = 0.8, importance = "impurity") %>%
-  # set_engine("ranger", num.threads = 8, replace = F, sample.fraction = 0.8, importance = "permutation", local.importance = T) %>%
-  # set_engine("ranger", num.threads = 8) %>%
-  fit(Exited ~ ., data = df_train)
+ranger_model_specs <- rand_forest("classification", 2, 1000, 5) %>% 
+  # set_engine("ranger", num.threads = 8, replace = F, sample.fraction = 0.8, importance = "impurity") %>%
+  set_engine("ranger", num.threads = 8, replace = F, sample.fraction = 0.8, importance = "permutation", local.importance = T)
+
+ranger_model_1 <- ranger_model_specs %>% fit(Exited ~ ., data = df_train1)
+
+ranger_model_2 <- ranger_model_specs %>% fit(Exited ~ ., data = df_train2)
 
-df_pred <- ranger_model_1 %>% 
-  predict(df_test) %>% 
-  bind_cols(df_test)
+df_pred1 <- ranger_model_1 %>% 
+  predict(df_test1) %>% 
+  bind_cols(df_test1)
+
+df_pred2 <- ranger_model_2 %>% 
+  predict(df_test2) %>% 
+  bind_cols(df_test2)
 
-df_pred %>% metrics(Exited, .pred_class)
+df_pred1 %>% metrics(Exited, .pred_class)
+df_pred2 %>% metrics(Exited, .pred_class)
 
-df_pred_probs <- ranger_model_1 %>% 
-  predict(df_test, type = "prob") %>% 
-  bind_cols(df_test)
+df_pred_probs1 <- ranger_model_1 %>% 
+  predict(df_test1, type = "prob") %>% 
+  bind_cols(df_test1)
 
-df_pred_probs %>% roc_auc(Exited, .pred_No)
-df_pred_probs %>% roc_curve(Exited, .pred_No) %>% autoplot()
+df_pred_probs2 <- ranger_model_2 %>% 
+  predict(df_test2, type = "prob") %>% 
+  bind_cols(df_test2)
+
+df_pred_probs1 %>% roc_auc(Exited, .pred_No)
+df_pred_probs2 %>% roc_auc(Exited, .pred_No)
+
+df_pred_probs1 %>% roc_curve(Exited, .pred_No) %>% autoplot()
+df_pred_probs2 %>% roc_curve(Exited, .pred_No) %>% autoplot()
 
 vi(ranger_model_1)
+vi(ranger_model_2)
+
 vip(ranger_model_1)
+vip(ranger_model_2)