使用 step_naomit 进行预测并使用 tidymodels 保留 ID

Predict with step_naomit and retain ID using tidymodels

在使用随机森林模型进行预测以合并回原始数据帧时,我试图在行中保留一个 ID。我在配方中使用 step_naomit,它在我烘焙训练数据时删除了缺少数据的行,但也删除了测试数据中缺少数据的记录。不幸的是,我没有 ID 无法轻松知道删除了哪些记录,因此我无法准确地合并回预测。

我曾尝试向原始数据中添加一个 ID 列,但 bake 会删除公式中未包含的任何变量(而且我不想在公式中包含 ID)。我还以为我可以保留原始 table 中的 row.names 进行合并,但看起来 row.name 在烘烤时也被重置了。

我意识到我可以在配方之前删除 NA 值来解决这个问题,但是配方中 step_naomit 的意义何在?我还在 step_naomit 中尝试了 skip=TRUE,但随后在拟合模型时出现缺少数据的错误(仅适用于随机森林)。我觉得我在 tidymodels 中遗漏了一些可以让我在烘焙前保留所有行的东西?

参见示例:


## R 3.6.1 ON WINDOWS 10 MACHINE

require(tidyverse)
require(tidymodels)
require(ranger)

set.seed(123)

temp <- iris %>%
    dplyr::mutate(Petal.Width = case_when(
        round(Sepal.Width) %% 2 == 0 ~ NA_real_, ## INTRODUCE NA VALUES
        TRUE ~ Petal.Width))

mySplit <- rsample::initial_split(temp, prop = 0.8)

myRecipe <- function(dataFrame) {
    recipes::recipe(Petal.Width ~ ., data = dataFrame) %>%
        step_naomit(all_numeric()) %>%
        prep(data = dataFrame)
}

myPred <- function(mySplit,myRecipe) {

    train_set <- training(mySplit)
    test_set <- testing(mySplit)

    train_prep <- myRecipe(train_set)

    analysis_processed <- bake(train_prep, new_data = train_set)

    model <- rand_forest(
            mode = "regression",
            mtry = 3,
            trees = 50) %>%
        set_engine("ranger", importance = 'impurity') %>%
        fit(Sepal.Width ~ ., data=analysis_processed)

    test_processed <- bake(train_prep, new_data = test_set)

    test_processed %>%
        bind_cols(myPrediction = unlist(predict(model,new_data=test_processed))) 

}

getPredictions <- myPred(mySplit,myRecipe)

nrow(getPredictions)

##  21 ROWS

max(as.numeric(row.names(getPredictions)))

##  21

nrow(testing(mySplit))

##  29 ROWS

max(as.numeric(row.names(testing(mySplit))))

##  150

为了能够跟踪删除了哪些观测值,我们需要为原始数据集提供一个 id 变量。

temp <- iris %>%
    dplyr::mutate(Petal.Width = case_when(
        round(Sepal.Width) %% 2 == 0 ~ NA_real_, ## INTRODUCE NA VALUES
        TRUE ~ Petal.Width),
        id = row_number()) #<<<<

然后我们使用 update_role() 首先将其指定为 "id variable",然后将其作为预测变量删除,这样它就不会成为建模过程的一部分。就是这样。其他一切都应该像以前一样工作。下面是完全更新的代码,其中包含 #<<<< 以表示我的更改。

require(tidyverse)
#> Loading required package: tidyverse
require(tidymodels)
#> Loading required package: tidymodels
#> Registered S3 method overwritten by 'xts':
#>   method     from
#>   as.zoo.xts zoo
#> ── Attaching packages ───────────────────── tidymodels 0.0.3 ──
#> ✔ broom     0.5.2     ✔ recipes   0.1.7
#> ✔ dials     0.0.3     ✔ rsample   0.0.5
#> ✔ infer     0.5.0     ✔ yardstick 0.0.4
#> ✔ parsnip   0.0.4
#> ── Conflicts ──────────────────────── tidymodels_conflicts() ──
#> ✖ scales::discard() masks purrr::discard()
#> ✖ dplyr::filter()   masks stats::filter()
#> ✖ recipes::fixed()  masks stringr::fixed()
#> ✖ dplyr::lag()      masks stats::lag()
#> ✖ dials::margin()   masks ggplot2::margin()
#> ✖ dials::offset()   masks stats::offset()
#> ✖ yardstick::spec() masks readr::spec()
#> ✖ recipes::step()   masks stats::step()
require(ranger)
#> Loading required package: ranger

set.seed(1234)

temp <- iris %>%
    dplyr::mutate(Petal.Width = case_when(
        round(Sepal.Width) %% 2 == 0 ~ NA_real_, ## INTRODUCE NA VALUES
        TRUE ~ Petal.Width),
        id = row_number()) #<<<<

mySplit <- rsample::initial_split(temp, prop = 0.8)

myRecipe <- function(dataFrame) {
    recipes::recipe(Petal.Width ~ ., data = dataFrame) %>%
        update_role(id, new_role = "id variable") %>%  #<<<<
        update_role(-id, new_role = 'predictor') %>%   #<<<<
        step_naomit(all_numeric()) %>%
        prep(data = dataFrame)
}

myPred <- function(mySplit,myRecipe) {

    train_set <- training(mySplit)
    test_set <- testing(mySplit)

    train_prep <- myRecipe(train_set)

    analysis_processed <- bake(train_prep, new_data = train_set)

    model <- rand_forest(
            mode = "regression",
            mtry = 3,
            trees = 50) %>%
        set_engine("ranger", importance = 'impurity') %>%
        fit(Sepal.Width ~ ., data=analysis_processed)

    test_processed <- bake(train_prep, new_data = test_set)

    test_processed %>%
        bind_cols(myPrediction = unlist(predict(model,new_data=test_processed))) 

}

getPredictions <- myPred(mySplit, myRecipe)

getPredictions
#> # A tibble: 23 x 7
#>    Sepal.Length Sepal.Width Petal.Length Petal.Width Species     id myPrediction
#>           <dbl>       <dbl>        <dbl>       <dbl> <fct>    <int>        <dbl>
#>  1          4.6         3.1          1.5         0.2 setosa       4         3.24
#>  2          4.3         3            1.1         0.1 setosa      14         3.04
#>  3          5.1         3.4          1.5         0.2 setosa      40         3.22
#>  4          5.9         3            4.2         1.5 versico…    62         2.98
#>  5          6.7         3.1          4.4         1.4 versico…    66         2.92
#>  6          6           2.9          4.5         1.5 versico…    79         3.03
#>  7          5.7         2.6          3.5         1   versico…    80         2.79
#>  8          6           2.7          5.1         1.6 versico…    84         3.12
#>  9          5.8         2.6          4           1.2 versico…    93         2.79
#> 10          6.2         2.9          4.3         1.3 versico…    98         2.88
#> # … with 13 more rows

# removed ids
setdiff(testing(mySplit)$id, getPredictions$id)
#> [1]   5  28  47  70  90 132

reprex package (v0.3.0)

于 2019-11-26 创建

step_naomit() 配方规范中使用 skip = TRUE,然后将配方包含在 workflow 中可能是正确的解决方案。例如,

myRecipe <- recipe(Petal.Width ~ ., data = dataFrame) %>%
        step_naomit(all_numeric(), step = FALSE)`
# don't include the prep()

wflow <- workflow() %>% 
  add_model(model) %>% 
  add_recipe(myRecipe)

wflow_fit <- wflow %>% 
  fit(train_set)

preds <- predict(wflow_fit, new_data = (test_set))