OneHotEncoder

OneHotEncoder is a preprocessing tool used in machine learning to convert categorical features into a numerical representation suitable for model training. For each categorical feature, it creates binary features (dummy variables) representing the presence of each possible category. For example, a feature color with values red, green, and blue becomes color_red, color_green, and color_blue. When multiple categorical features are encoded, their binary indicators are concatenated to form a larger feature set.

The encoder can output a sparse matrix, which is memory-efficient when there are many categories or features with many possible values. Depending on the configuration, it can also produce a dense array. A common option is to drop one category per feature to reduce dimensionality and avoid multicollinearity in certain models, controlled by a drop parameter (for example drop='first').

In practice, OneHotEncoder is often used within machine learning pipelines. It can handle string or numeric categories, infer category levels automatically, and transform data after fitting on training data. Some implementations support handling unknown categories encountered during transformation, typically via a handle_unknown option set to ignore or error. It is commonly integrated with tools like scikit-learn’s ColumnTransformer to apply encoding to selected columns while leaving others untouched, and it can be configured to emit sparse or dense output.

Limitations include a rapid increase in dimensionality when features have many categories, which can lead to sparsity and increased computation. This makes one-hot encoding less suitable for high-cardinality features without additional strategies. Alternatives such as ordinal encoding, target encoding, or embeddings may be preferred in some scenarios. OneHotEncoder is widely used for preparing data for linear models and tree-based methods that require numerical input.