Summary of "Encoding Categorical Data | Ordinal Encoding | Label Encoding"

Encoding Categorical Data (Ordinal Encoding, Label Encoding)

Main ideas and concepts

Categorical features must be converted to numeric values before using most machine learning algorithms.
Two main categorical types:
- Nominal: categories have no order or relationship (e.g., state, gender, product category).
- Ordinal: categories have a meaningful order (e.g., education levels, review ratings).
Appropriate encoding methods:
- Use OrdinalEncoder (or a custom mapping) for ordinal features.
- Use one-hot encoding for nominal features (covered in a separate video).
- LabelEncoder is intended for target labels (y) in classification and is not generally appropriate for input features.

Important practical rules

Always split data into train and test before fitting encoders to avoid data leakage. Fit encoders on the training set, then transform both train and test with the fitted encoder.

If you don’t explicitly tell OrdinalEncoder the category order, numeric labels may be assigned arbitrarily.
Use ColumnTransformer and Pipeline to apply different transformations to different columns in a clean, maintainable way.

Step-by-step methodology / instructions

Identify categorical columns and their types
- Determine which features are nominal vs ordinal using domain knowledge.
- Example dataset columns:
  - gender (nominal)
  - reviews (ordinal: poor / average / good)
  - education (ordinal: school / undergraduate / postgraduate)
  - purchased (yes/no — nominal target)
Split data into train and test
- Use train_test_split (or equivalent).
- Always learn encoders only on the training set.
Prepare OrdinalEncoder for ordinal features
- Import OrdinalEncoder from sklearn.preprocessing (or implement a custom mapping).
- When constructing the encoder, pass a categories parameter that explicitly specifies the ordered categories for each ordinal column (a list of lists), for example:
  - reviews: ['poor', 'average', 'good']
  - education: ['school', 'undergraduate', 'postgraduate']
- Fit the encoder on X_train and transform both X_train and X_test.
- Result: ordinal categories convert to integers reflecting the given order (e.g., poor → 0, average → 1, good → 2).
Handle nominal features
- Use OneHotEncoder (or another appropriate encoder) for nominal features to avoid introducing artificial order.
- Apply OneHotEncoder via ColumnTransformer or a pipeline for selected nominal columns.
Encode the target variable (if needed)
- Use LabelEncoder only for target labels (y) in classification tasks.
- Fit LabelEncoder on y_train and transform y_train and y_test.
- Do not use LabelEncoder for input features unless you intentionally want a numeric mapping and understand the implications.
Combine transformations cleanly
- Use ColumnTransformer to assign different transformers (OrdinalEncoder, OneHotEncoder, etc.) to different subsets of columns.
- Use Pipeline to chain preprocessing and modeling steps and ensure reproducible transforms.

Examples & warnings highlighted

Example mappings:
- reviews (poor < average < good) → ordinal numbers.
- education (school < undergraduate < postgraduate) → ordinal numbers.
If categories/order aren’t supplied to OrdinalEncoder, category-to-number assignments can be arbitrary and inconsistent.
Historically some practitioners used LabelEncoder on input features; this is discouraged because it can imply an order that doesn’t exist.
Always fit transformers on training data only, then apply to the test set.

Tools / classes mentioned

sklearn.preprocessing.OrdinalEncoder
sklearn.preprocessing.LabelEncoder
sklearn.model_selection.train_test_split
ColumnTransformer
Pipeline
OneHotEncoder

Speakers / sources featured

Video presenter / YouTuber (unnamed) — sole speaker explaining concepts and demonstrating code examples.
Referenced documentation/tools: scikit-learn classes (OrdinalEncoder, LabelEncoder, train_test_split, ColumnTransformer, OneHotEncoder).