欠損値に基づく選択

Rによる次元削減

Matt Pickard

Owner, Pickard Predictives, LLC

欠損値比を計算する

欠損値比の式

n <- nrow(credit_df)


missing_vals_df <- credit_df %>% 
  summarize(across(everything(), ~ sum(is.na(.)))) %>%

  pivot_longer(everything(), names_to = "feature", values_to = "num_missing_values") %>%

  mutate(missing_val_ratio = num_missing_values / n)

欠損値比の出力

missing_vals_df

# A tibble: 5 × 3
  feature          num_missing_values missing_val_ratio
  <chr>                         <int>             <dbl>
1 credit_score                      0             0    
2 annual_income                     0             0    
3 age                              84             0.613
4 outstanding_debt                129             0.942
5 num_of_loan                       0             0

欠損値比しきい値の目安

客観的なカットオフはない
特徴量の重要度に依存
- 例: outstanding_debt と age

欠損値比しきい値の目安表

欠損値フィルタを作成

missing_vals_filter <- missing_vals_df %>% 
  filter(missing_val_ratio <= 0.5) %>%

  pull(feature)


missing_vals_filter

[1] "credit_score"  "annual_income" "num_of_loan"

欠損値フィルタの適用

filtered_credit_df <- credit_df %>% 
  select(missing_vals_filter)


filtered_credit_df %>% head(3)

# A tibble: 5 × 3
  credit_score annual_income num_of_loan
  <chr>                <dbl>       <dbl>
1 Standard            87630.           4
2 Standard            16574.           7
3 Standard            24931.           2

tidymodels の方法

レシピを作成

missing_vals_recipe <- 
  recipe(credit_score ~ ., data = credit_df) %>%

  step_filter_missing(all_predictors(), threshold = 0.5) %>% 

  prep()

レシピを適用

filtered_credit_df <- 
  bake(missing_vals_recipe, new_data = NULL)

ベイク後のレシピ出力

filtered_credit_df %>% head(5)

# A tibble: 5 × 3
  annual_income num_of_loan credit_score
          <dbl>       <dbl> <fct>       
1        87630.           4 Standard    
2        16574.           7 Standard    
3        24931.           2 Standard    
4       136680.           1 Good        
5        76850.           3 Standard

Passons à la pratique !

Rによる次元削減