63

Tidymodels 入门

现代建模框架

Tidymodels 篇

Rstudio工厂的 Max Kuhn 大神正主持机器学习的开发，日臻成熟了，感觉很强大啊。

R 
library(tidyverse)
library(tidymodels)

数据

R 
penguins <- read_csv("./demo_data/penguins.csv") %>%
  janitor::clean_names() %>% 
  drop_na()

penguins %>%
  head()
R 
penguins %>%
  ggplot(aes(x = bill_length_mm, y = bill_depth_mm, 
             color = species, shape = species)
         ) +
  geom_point()

机器学习

R 
split <- penguins %>% 
  mutate(species = as_factor(species)) %>% 
  mutate(species = fct_lump(species, 1)) %>% 
  initial_split()

split

training_data <- training(split)
training_data
testing_data <- testing(split)
testing_data

model01

R 
model_logistic <- parsnip::logistic_reg() %>% 
  set_engine("glm") %>% 
  set_mode("classification") %>% 
  fit(species ~ bill_length_mm + bill_depth_mm, data = training_data)


bind_cols(
  predict(model_logistic, new_data = testing_data, type = "class"),
  predict(model_logistic, new_data = testing_data, type = "prob"),
  testing_data
)


predict(model_logistic, new_data = testing_data) %>% 
  bind_cols(testing_data) %>% 
  count(.pred_class, species)

model02

R 
model_neighbor <- parsnip::nearest_neighbor(neighbors = 10) %>% 
  set_engine("kknn") %>% 
  set_mode("classification") %>% 
  fit(species ~ bill_length_mm, data = training_data)

predict(model_neighbor, new_data = testing_data) %>% 
  bind_cols(testing_data) %>% 
  count(.pred_class, species)

model03

R 
model_multinom <- parsnip::multinom_reg() %>% 
  set_engine("nnet") %>% 
  set_mode("classification") %>% 
  fit(species ~ bill_length_mm, data = training_data)

predict(model_multinom, new_data = testing_data) %>% 
  bind_cols(testing_data) %>% 
  count(.pred_class, species)

model04

R 
model_decision <- parsnip::decision_tree() %>% 
  set_engine("rpart") %>% 
  set_mode("classification") %>% 
  fit(species ~ bill_length_mm, data = training_data)

predict(model_decision, new_data = testing_data) %>% 
  bind_cols(testing_data) %>% 
  count(.pred_class, species)

workflow

使用 recipes

R 
library(tidyverse)
library(tidymodels)
library(workflows)

penguins <- readr::read_csv("./demo_data/penguins.csv") %>%
  janitor::clean_names() 


split <- penguins %>% 
  tidyr::drop_na() %>% 
  rsample::initial_split(prop = 3/4)

training_data <- rsample::training(split)
testing_data  <- rsample::testing(split)

参考tidy modeling in R, 被预测变量在分割前，应该先处理，比如标准化。 但这里的案例，我为了偷懒，被预测变量bill_length_mm，暂时保留不变。 预测变量做标准处理。

R 
penguins_lm <- 
  parsnip::linear_reg() %>% 
  #parsnip::set_engine("lm") 
  parsnip::set_engine("stan") 

penguins_recipe  <- 
  recipes::recipe(bill_length_mm ~ bill_depth_mm + sex, data = training_data) %>% 
  recipes::step_normalize(all_numeric(), -all_outcomes()) %>% 
  recipes::step_dummy(all_nominal())


broom::tidy(penguins_recipe)
图片
R 
penguins_recipe %>% 
  recipes::prep(data = training_data) %>%  #or prep(retain = TRUE)
  recipes::juice()


penguins_recipe %>%   
  recipes::prep(data = training_data) %>% 
  recipes::bake(new_data = testing_data)   # recipe used in new_data



train_data <- 
  penguins_recipe %>%   
  recipes::prep(data = training_data) %>% 
  recipes::bake(new_data = NULL) 


test_data <- 
  penguins_recipe %>%   
  recipes::prep(data = training_data) %>% 
  recipes::bake(new_data = testing_data)

workflows的思路更清晰

workflows的思路让模型结构更清晰。 这样prep(), bake(), and juice() 就可以省略了，只需要recipe和model，他们往往是成对出现的

R 
wflow <- 
  workflows::workflow() %>% 
  workflows::add_recipe(penguins_recipe) %>% 
  workflows::add_model(penguins_lm) 


wflow_fit <- 
  wflow %>% 
  parsnip::fit(data = training_data)
R 
wflow_fit %>% 
  workflows::pull_workflow_fit() %>% 
  broom.mixed::tidy()

wflow_fit %>% 
  workflows::pull_workflow_prepped_recipe()

先提取模型，用在 predict() 是可以的，但这样太麻烦了

R 
wflow_fit %>% 
  workflows::pull_workflow_fit() %>% 
  stats::predict(new_data = test_data) # note: test_data not testing_data

因为，predict() 会自动的将recipes(对training_data的操作)，应用到testing_data 这个不错，参考这里

R 
penguins_pred <- 
  predict(
    wflow_fit, 
    new_data = testing_data %>% dplyr::select(-bill_length_mm), # note: testing_data not test_data
    type = "numeric"
  ) %>% 
  dplyr::bind_cols(testing_data %>% dplyr::select(bill_length_mm))

penguins_pred
R 
penguins_pred %>% 
  ggplot(aes(x = bill_length_mm, y = .pred)) + 
  geom_abline(linetype = 2) + 
  geom_point(alpha = 0.5) + 
  labs(y = "Predicted ", x = "bill_length_mm")

augment()具有predict()一样的功能和特性，还更简练的多

R 
wflow_fit %>%
  augment(new_data = testing_data) %>%       # note: testing_data not test_data
  ggplot(aes(x = bill_length_mm, y = .pred)) + 
  geom_abline(linetype = 2) + 
  geom_point(alpha = 0.5) + 
  labs(y = "Predicted ", x = "bill_length_mm")

模型评估

参考<https://www.tmwr.org/performance.html#regression-metrics>

R 
penguins_pred %>%
  yardstick::rmse(truth = bill_length_mm, estimate = .pred)

自定义一个指标评价函数my_multi_metric，就是放一起，感觉不够tidyverse

R 
my_multi_metric <- yardstick::metric_set(rmse, rsq, mae, ccc)

penguins_pred %>%
  my_multi_metric(truth = bill_length_mm, estimate = .pred)
R 
# remove the objects
# ls() %>% stringr::str_flatten(collapse = ", ")

rm(my_multi_metric, penguins, penguins_lm, penguins_pred, penguins_recipe, split, testing_data, training_data, wflow, wflow_fit)
R 
pacman::p_unload(pacman::p_loaded(), character.only = TRUE)