LabelEncoding

LabelEncoding is a data preprocessing technique that assigns each distinct category in a categorical variable a unique integer. It is commonly used to convert the target variable in supervised learning tasks, and, less commonly, to transform categorical features when the machine learning model can interpret ordinal relationships.

How it works: Each category is mapped to a numeric label, typically starting at 0. For example, categories ['cat','dog','mouse'] might be encoded as {'cat': 0, 'dog': 1, 'mouse': 2}, producing a numeric array such as [0, 1, 0, 2]. In practice, most implementations determine the mapping during a fit stage and then apply it with transform.

Usage and limitations: Label encoding is most appropriate for the target variable in classification or regression problems. It is generally not ideal for nominal features because it imposes an arbitrary order, which can mislead models that assume ordinal relationships (e.g., linear models). For algorithms that can handle ordinal input or when used with tree-based models, it may be acceptable for features in some cases. A key limitation is that, in many tools, the encoder cannot gracefully handle unseen categories in new data unless explicit handling is added.

Implementation notes: In scikit-learn, LabelEncoder operates on 1D arrays and provides fit, transform, fit_transform, and inverse_transform. It should be fitted on the training data and then applied to validation or test data to ensure consistent labeling. If a new category appears during transform, some implementations raise an error.

Alternatives: One-Hot Encoding (OneHotEncoder) avoids implying ordinal relationships; other approaches include target encoding or binary encoding, depending on the task and model.