---
title: "Stat. 652 - NHANES RandomForest"
output: html_notebook
---

Here we try using Tidymodels and Recipe to fit a classifier, the RandomForest model.

Note the need to impute values because of the high percentage of missing values.

For the RandomForest model we use all of the features in the data set.

```{r}
library(pacman)

p_load(NHANES, tidyverse, tidymodels, naniar)
```

### Step 1: Access the data.

```{r}
data(NHANES)

NHANES
```

### Step 2: Clean the data.  Prepare the data and make a recipe for applying the steps needed to preprocess the data.

First drop all of the rows where the y-variable SleepTrouble is missing.

```{r}
NHANES_SleepTrouble <- NHANES %>% select(-ID, -SleepHrsNight) %>%
  select( SleepTrouble, everything()) %>%
  drop_na(SleepTrouble)
NHANES_SleepTrouble

```

Summarize the y-variable.

```{r}
NHANES_SleepTrouble %>% group_by(SleepTrouble) %>%
  summarize(n = n()) %>%
  mutate(freq = n / sum(n))
```

Make the first split with 80% of the data being in the trainning data set.

```{r}
NHANES_SleepTrouble_split <- initial_split(NHANES_SleepTrouble, prop = 0.8)
NHANES_SleepTrouble_split
```

Trainning data.

```{r}
NHANES_SleepTrouble_split %>%
  training() 
```

```{r}
NHANES_SleepTrouble_split %>%
  training() %>%
  vis_miss()
```

Create the recipe for applying the preprocessing.  Note the use of step_nzv(), which removes any columns that have very low variability, and the use of the step_knnimpute() function, which fills in the cells that are missing with the median of the column.

```{r}
NHANES_SleepTrouble_recipe <- training(NHANES_SleepTrouble_split) %>%
  recipe(SleepTrouble ~ .) %>%
  step_nzv(all_predictors()) %>%
  step_knnimpute(all_predictors()) %>%
  prep()

summary(NHANES_SleepTrouble_recipe)

tidy(NHANES_SleepTrouble_recipe)
```


```{r}
NHANES_SleepTrouble_testing <- NHANES_SleepTrouble_recipe %>%
  bake(testing(NHANES_SleepTrouble_split)) 

NHANES_SleepTrouble_testing
```


```{r}
NHANES_SleepTrouble_training <- juice(NHANES_SleepTrouble_recipe)

NHANES_SleepTrouble_training
```


### Step 3: Training a model on the data

Setup the models.

```{r}
NHANES_SleepTrouble_ranger <- rand_forest(trees = 100) %>%
  set_engine("ranger") %>%
  set_mode("classification") %>%
  fit(SleepTrouble ~ ., data = NHANES_SleepTrouble_training)
```


```{r}
predict(NHANES_SleepTrouble_ranger, NHANES_SleepTrouble_testing)
```

```{r}
NHANES_SleepTrouble_ranger %>%
  predict(NHANES_SleepTrouble_testing) %>%
  bind_cols(NHANES_SleepTrouble_testing) 
```

### Step 4: Evaluate the models.

```{r}
NHANES_SleepTrouble_ranger %>%
  predict(NHANES_SleepTrouble_testing) %>%
  bind_cols(NHANES_SleepTrouble_testing) %>%
  metrics(truth = SleepTrouble, estimate = .pred_class)
```


```{r}
NHANES_SleepTrouble_ranger %>%
  predict(NHANES_SleepTrouble_testing) %>%
  bind_cols(NHANES_SleepTrouble_testing) %>%
  conf_mat(truth = SleepTrouble, estimate = .pred_class)

```


```{r}
NHANES_SleepTrouble_ranger %>%
  predict(NHANES_SleepTrouble_testing, type = "prob") %>%
  bind_cols(NHANES_SleepTrouble_testing) %>%
  roc_curve(SleepTrouble, .pred_Yes) %>%
  ggplot(aes(x = 1 - specificity, y = sensitivity)) +
  geom_path() +
  geom_abline(lty = 3) +
  coord_equal() 
```