特徵選取 vs. 特徵擷取

R 的降維

Matt Pickard

Owner, Pickard Predictives, LLC

降維方法

特徵選取 像拔雜草
特徵擷取 像做沙拉

¹ 圖片來源：Daderot，CC0，取自 Wikimedia Commons

特徵選取

六個以色彩標示的特徵集

特徵選取

篩掉低資訊特徵的六個特徵集

特徵選取

篩選後的四個特徵

信用資料範例

credit_df %>% head(n=5)

  annual_income num_bank_accounts num_credit_card outstanding_debt credit_history_months
          <dbl>             <dbl>           <dbl>            <dbl>                 <dbl>
1        87630.                 2               5             526.                   286
2        16574.                 2               5              NA                    122
3        24931.                 2               5              NA                    351
4       136680.                 2               5              NA                    216
5        76850.                 2               5            1112.                   272

建立零變異篩選器

na_filter <- credit_df %>% 
  summarize(across(everything(), ~ var(., na.rm = TRUE))) %>%

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

  filter(variance == 0) %>%

  pull(feature)

na_filter

"num_bank_accounts" "num_credit_card"

建立遺漏值篩選器

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

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

  filter(num_missing_values > 0) %>%

  pull(feature)

na_filter

"outstanding_debt"

套用合併篩選器

combined_filter <- 
  c(low_var_filter, na_filter)

credit_df %>% 
  select(-all_of(combined_filter)) %>% 
  head(3)

  annual_income credit_history_months
          <dbl>                 <dbl>
1        87630.                   286
2        16574.                   122
3        24931.                   351