credit_scoring.R

library(dplyr)
library(DALEX)
library(skimr)
library(forcats)
library(randomNames)

raw <- read.csv("./CreditScoring.csv")

raw %>%
  filter(Status != 0, Home !=0, Marital !=0, Job != 0) %>%
  filter_all(function(x) x != 99999999) %>%
  mutate(
    Status = fct_recode(as.factor(Status), negative="2", positive="1") %>% fct_rev,
    Home = fct_recode(as.factor(Home), rent="1", notarial_act="2", private_contract="3", ignoring_contract="4", parents="5", other="6"),
    Marital = fct_recode(as.factor(Marital), single="1", married="2", widow="3", separated="4", divorced="5"),
    Records = fct_recode(as.factor(Records), no="1", yes="2"),
    Job = fct_recode(as.factor(Job), fixed="1", partime="2", freelance="3", other="4")
  ) -> df


set.seed(1313)
rnd_names <- randomNames(n=nrow(df)*3, name.order="first.last", name.sep=" ") %>% unique
rownames(df) <- rnd_names[seq_len(nrow(df))]

train_index <- sample(1:nrow(df), size=round(0.6 * nrow(df)))
train <- df[train_index, ]
test <- df[-train_index, ]

library(arenar)
library(randomForest)
model <- randomForest(Status ~ ., data = train)
explainer <- DALEX::explain(model, data = test[,-1], y = test$Status == "positive")

model_performance(explainer)
# Measures for:  classification
# recall     : 0.9128728 
# precision  : 0.8161404 
# f1         : 0.8618007 
# accuracy   : 0.7868571 
# auc        : 0.8287271
# 
# Residuals:
#      0%     10%     20%     30%     40%     50%     60%     70%     80% 
# -0.9920 -0.6122 -0.3640  0.0080  0.0340  0.0660  0.1120  0.1840  0.2800 
#     90%    100% 
#  0.4300  0.8640 

             
create_arena(live=T) %>% push_model(explainer) %>% push_observations(test) %>% push_observations(changed) %>% run_server(port=9763)