330_featureSelectionExample.qmd

# Feature Selection Example

This section presents a modern feature-selection workflow using `tidymodels`.
The goal is to keep the process explicit and reproducible:

- split data into training and test subsets
- define feature filters in a `recipe`
- fit a model through a `workflow`
- inspect variable importance

```{r setup_fs, message=FALSE, warning=FALSE}
library(tidymodels)

has_vip <- requireNamespace("vip", quietly = TRUE)

set.seed(10)
kc1 <- read.csv("./datasets/defectPred/unified/Unified-file.csv", stringsAsFactors = FALSE)
kc1 <- kc1[, c("McCC", "CLOC", "PDA", "PUA", "LLOC", "LOC", "bug")]
kc1$Defective <- factor(ifelse(kc1$bug > 0, "Y", "N"))
kc1$bug <- NULL

split <- initial_split(kc1, prop = 0.75, strata = Defective)
training <- training(split)
testing <- testing(split)
```

## Feature filtering with recipes

`recipes` supports common feature-selection and cleaning steps:

- `step_zv()` removes zero-variance predictors
- `step_nzv()` removes near-zero-variance predictors
- `step_corr()` removes highly correlated numeric predictors

```{r fs_recipe, message=FALSE, warning=FALSE}
rec <- recipe(Defective ~ ., data = training) |>
  step_zv(all_predictors()) |>
  step_nzv(all_predictors()) |>
  step_corr(all_numeric_predictors(), threshold = 0.90)

rec_prep <- prep(rec)
juice(rec_prep) |> dplyr::glimpse()
```

## Leakage pitfalls in feature selection

Feature selection must be fit on training data only. Common mistakes:

- selecting features on the full dataset before splitting
- ranking features using labels from future releases
- including attributes that are proxies of the target (post-release fields)

In this chapter, feature filters are defined in the recipe and learned from
`training` only, then applied to `testing` through the fitted workflow.

## Train a model after feature filtering

```{r fs_model, message=FALSE, warning=FALSE}
rf_spec <- rand_forest(trees = 500, min_n = 5) |>
  set_mode("classification") |>
  set_engine("ranger", importance = "impurity")

wf <- workflow() |>
  add_recipe(rec) |>
  add_model(rf_spec)

rf_fit <- fit(wf, data = training)
rf_fit
```

## Evaluate on the test set

```{r fs_eval, message=FALSE, warning=FALSE}
pred_cls <- predict(rf_fit, testing, type = "class")
pred_prb <- predict(rf_fit, testing, type = "prob")

eval_tbl <- bind_cols(testing, pred_cls, pred_prb)

metrics(eval_tbl, truth = Defective, estimate = .pred_class)
conf_mat(eval_tbl, truth = Defective, estimate = .pred_class)
```

## Variable importance

```{r fs_importance, message=FALSE, warning=FALSE, fig.width=8, fig.height=6}
if (has_vip) {
  rf_fit |>
    extract_fit_parsnip() |>
    vip::vip(num_features = 15)
} else {
  message("Package 'vip' is not installed; skipping variable-importance plot.")
}
```

## Feature Selection Packages and Further Reading

For this chapter, the most useful package families are:

| Package | Typical Use in Feature Selection |
|---------|----------------------------------|
| [recipes](https://recipes.tidymodels.org/) | Filtering and preprocessing (`step_zv`, `step_nzv`, `step_corr`) |
| [Boruta](https://cran.r-project.org/package=Boruta) | All-relevant feature selection using random forests |
| [FSelectorRcpp](https://cran.r-project.org/package=FSelectorRcpp) | Information-gain and entropy-based ranking |
| [vip](https://koalaverse.github.io/vip/) | Variable-importance visualization for fitted models |
| [ranger](https://cran.r-project.org/package=ranger) | Fast tree-based models with built-in importance scores |

To keep this chapter concise, we focus on the `recipes` + `ranger` workflow. See @sec-popular-packages for a broader package map and @sec-rpackages for package-management guidance.