Hackaton

library(tidyverse)
library(tidymodels)
library(textrecipes)
library(stopwords)

hotels <- read_csv("./tripadvisor_hotel_reviews.csv")
set.seed(1)
hotels <- sample_frac(hotels, 0.3)

set.seed(1)
split <- hotels |> 
  initial_split()

train <- training(split)
test <- testing(split)

svm_spec <- svm_linear() |>
  set_engine("LiblineaR") |> 
  set_mode("regression")
#добавим модель к воркфлоу
svm_wflow <- workflow() |> 
  add_model(svm_spec)

svm_wflow
#напишем рецепт препросессинга данных
rec <- recipe(Rating ~ Review, data = train) |> 
  step_tokenize(Review)  |> 
  step_tokenfilter(Review, max_tokens = 1000)  |> 
  step_tfidf(Review) 


#Добавляем препроцессор в воркфлоу.
svm_wflow <- svm_wflow |> 
  add_recipe(rec)

svm_wflow

#подгонка модели на обучающих данных
svm_fit <- svm_wflow |>
  fit(data = train)
  
#Оценим модель на тренировочных данных.
pred_data <- tibble(truth = train$Rating,
                    estimate = predict(svm_fit, train)$.pred)
pred_data
metrics <- metric_set(rmse, rsq, mae)

metrics_svm <- metrics(pred_data, truth = truth,  estimate = estimate) 

1 rmse standard 1.1387260 2 rsq standard 0.4952173 3 mae standard 0.9003342

#заменим выборку
set.seed(1)
folds <- vfold_cv(train, v = 10) 

#нулевая модель
null_reg <- null_model() |> 
  set_engine("parsnip") |> 
  set_mode("regression")

null_wflow <- workflow() |> 
  add_model(null_reg) |> 
  add_recipe(rec)

null_rs <- fit_resamples(
  null_wflow,
  folds,
  control = control_resamples(save_pred = TRUE)
)
collect_metrics(null_rs)

1 rmse standard 1.237592 10 0.0111641 Preprocessor1_Model1 2 rsq standard NaN 0 NA Preprocessor1_Model1

install.packages("spacyr")
spacyr::spacy_install()

rec_lemma <- recipe(Rating ~ Review, data = train) |> 
  step_tokenize(Review, engine = "spacyr")  |> 
  step_lemma(Review) |>
  step_tokenfilter(Review, max_tokens = 1000)  |> 
  step_tfidf(Review) 
svm_wflow_lemma <- workflow() |> 
  add_model(svm_spec)|> 
  add_recipe(rec_lemma)

svm_fit_lemma <- svm_wflow_lemma |>
  fit(data = train)

pred_data_lemma <- tibble(truth = train$Rating,
                    estimate = predict(svm_fit_lemma, train)$.pred)
pred_data_lemma
metrics_lemma <- metric_set(rmse, rsq, mae)

metrics_svm_lemma <- metrics(pred_data_lemma, truth = truth,  estimate = estimate) 

1 rmse standard 1.1569035 2 rsq standard 0.4700319 3 mae standard 0.9135093

stopwords_rec <- function(stopwords_name) {
  recipe(Rating ~ Review, data = train) |> 
  step_tokenize(Review)  |> 
  step_stopwords(Review, stopword_source = stopwords_name) |> 
  step_tokenfilter(Review, max_tokens = 1000)  |> 
  step_tfidf(Review) 
}
set.seed(1)
svm_wflow_stop <- workflow() |> 
  add_model(svm_spec)

snowball_rs <-  fit_resamples(
  svm_wflow_stop |>  
  add_recipe(stopwords_rec("snowball")),
  folds
)
metrics_snowball <- collect_metrics(snowball_rs)

1 rmse standard 1.1471342 10 0.009814347 Preprocessor1_Model1 2 rsq standard 0.4621366 10 0.008786747 Preprocessor1_Model1

set.seed(1)
smart_rs <- fit_resamples(
  svm_wflow_stop |> add_recipe(stopwords_rec("smart")),
  folds
)
metrics_smart_rs <- collect_metrics(smart_rs) 

1 rmse standard 1.1404006 10 0.009630983 Preprocessor1_Model1 2 rsq standard 0.4668183 10 0.009518076 Preprocessor1_Model1

set.seed(1)
stopwords_iso_rs <- fit_resamples(
  svm_wflow_stop |> add_recipe(stopwords_rec("stopwords-iso")),
  folds
)
metrics_stopwords_iso_rs <- collect_metrics(stopwords_iso_rs) 

1 rmse standard 1.1324068 10 0.009622164 Preprocessor1_Model1 2 rsq standard 0.4546836 10 0.009070150 Preprocessor1_Model1

xgb_spec <- boost_tree(mtry = 50, trees = 500)  |> 
  set_engine("xgboost")  |> 
  set_mode("regression")

xgb_wflow <- workflow() |> 
  add_model(xgb_spec) |> 
  add_recipe(rec)

xgb_wflow

xgb_rs <- fit_resamples(
  xgb_wflow,
  folds,
  control = control_resamples(save_pred = TRUE)
)

1 rmse standard 0.8439222 10 0.010151510 Preprocessor1_Model1 2 rsq standard 0.5370892 10 0.009386127 Preprocessor1_Model1

# ОБУЧЕНИЕ ФИНАЛЬНОЙ МОДЕЛИ НА ВСЕХ ТРЕНИРОВОЧНЫХ ДАННЫХ
final_fit <- last_fit(xgb_wflow, split)

# Метрики на тестовых данных
test_metrics <- collect_metrics(final_fit)

1 rmse standard 0.8570846 Preprocessor1_Model1 2 rsq standard 0.5139780 Preprocessor1_Model1