Chapter 5 Classification I: training & predicting

When should you use classification?

When the response variable is categorical (e.g., yes/no, species type) and you want to predict class labels.

What is a training dataset?

A dataset with predictors and known labels used to teach the model patterns for making predictions on new data

What is Euclidean distance?

The straight-line distance between two points in space.

What is K-nearest neighbors (KNN)?

A classification algorithm that assigns a class based on the majority label among the K closest training points.

Why is scaling important in KNN?

Because distance calculations are sensitive to variable scale; larger-scale variables dominate distances.

Steps in KNN classification workflow

1. Split data (train/test)

2. Preprocess (scale, center, clean)

3. Specify model

4. Fit model to training data

5. Predict on new/test data

6. Evaluate performance

What is a recipe in tidymodels?

A set of preprocessing steps applied to data before modeling.

It includes steps such as centering and scaling predictors, handling missing values (imputation), and balancing classes (e.g., upsampling).

Centering vs scaling

• Centering: subtract mean

• Scaling: divide by standard deviation

step_center(all_predictors()) + step_scale(all_predictors())

What is imputation?

Filling in missing values.

Syntax: step_impute_mean(all_predictors())

What is balancing data?

Adjusting class distribution to avoid bias toward majority class.

Syntax to upsample: step_upsample(outcome_variable)

What does bake() do?

Applies the trained recipe to new data

KNN model specification in R

knn_spec <- nearest_neighbor(weight_func = "rectangular", neighbors = 5) |>
set_engine("kknn") |>
set_mode("classification")

What does set_mode("classification") do?

Tells the model to predict categorical outcomes.

What is a workflow and why do we use it?

Combines preprocessing (recipe) and model into one object. Ensures consistent application of preprocessing and modeling steps.

glimpse()

Quickly view structure of a dataset

distinct()

Returns unique rows/values — removes duplicate rows

fct_recode()

Is used to rename or relabel the categories (levels) of a factor variable.

For example, if a variable has levels like "M" and "F", you can change them to "Male" and "Female" without altering the data itself.

drop_na()

Is used in a data wrangling pipeline to remove rows that contain missing (NA) values

data |> drop_na()

bind_rows() vs bind_cols()

• bind_rows(): stacks datasets vertically

• bind_cols(): combines datasets horizontally

When to use drop_na() vs imputation?

• drop_na(): when few missing values

• imputation: when you want to keep data

drop_na() removes entire rows that contain missing values, so you lose those observations.

Imputation fills in missing values (e.g., using the mean), allowing you to keep all observations.

Why and how preprocess data?

• Ensure variables are comparable and usable

Common steps:

recipe(class ~ ., data = training_data) |>
  step_scale(all_predictors()) |>
  step_center(all_predictors()) |>
  step_impute_mean(all_predictors()) |>
  step_downsample(class)