kaggle.R

In alastairrushworth/mlblitz: ML: getting it done.

library(tidyverse)
devtools::load_all()



# load in the train and test data
load("/Users/alastairrushworth/Documents/git_repositories/mlblitz/data/fraud/train_test.RData")

train <- prepare(train_test[[1]], target = "isFraud")
test  <- prepare(train_test[[2]], train, target = "isFraud")





# train %>% select_if(is.character) %>% inspect_cat %>% show_plot

# ensure recipe can be extended with custom functions
# test_blind option?... if want to borrow feature information from test
# comparison of difference in train and test
# feature generation using grouped summaries
# conversion of NAs to binary
# save recipes for repeating on test set - add attribute
train <- train %>% mutate_if(is.character, as.factor)
test  <- test  %>% mutate_if(is.character, as.factor)

# 
# train %>% inspect_types
# 
# z <- train %>% select_if(is.logical)
# z$isFraud <- train$isFraud
# z <- z %>% mutate_if(is.logical, as.character) 
# z[is.na(z)] <- "missing"
# dfs <- z %>% group_by(isFraud) %>% nest %>% .$data
# x <- inspect_cat(dfs[[1]], dfs[[2]])



# devtools::load_all("/Users/alastairrushworth/Documents/git_repositories/inspectdf")

# get the label out
label  <- train["isFraud"] %>% 
  mutate(isFraud = as.factor(make.names(isFraud))) %>%
  .$isFraud
train_d <- train %>% select(-isFraud) 


fit_params <- list(iterations = 200,
                   loss_function = 'Logloss',
                   border_count = 32,
                   depth = 5,
                   learning_rate = 0.03,
                   l2_leaf_reg = 3.5)

library(caret)
library(catboost)

fit_control <- trainControl(method = "cv",
                            number = 2,
                            classProbs = TRUE)

grid <- expand.grid(depth = c(6, 8),
                    learning_rate = 0.1,
                    iterations = 100,
                    l2_leaf_reg = 1e-3,
                    rsm = 0.95,
                    border_count = 64)

report <- train(train_d, label,
                method = catboost.caret,
                logging_level = 'Verbose', preProc = NULL,
                tuneGrid = grid, trControl = fit_control)

print(report)
importance <- varImp(report, scale = FALSE)
print(importance)



train_pool  <- catboost.load_pool(train_d, label = as.integer(unlist(label)))
model <- catboost.train(pool, params = fit_params)

test_pool <- catboost.load_pool(test)
preds <- catboost.predict(model, test_pool, prediction_type = 'Probability')

# read in the sample submission file
ss <- read_csv("data/fraud/sample_submission.csv")
ss$isFraud <- preds
write_csv(ss, "data/fraud/sub.csv")