Machine learning models are used in many applications to predict a target variable given features that provides additional information. For instance, we can be interested in predicting the salary of an employee (target) given its job title and experience level (features). However, machine learning algorithms only operate on numerical features. When dealing with categorical features like job titles, a preprocessing step must be performed in order to encode them into numerical features.

A naive way to encode such data into numerical features is one-hot encoding: each job title is assigned a binary vector that is orthogonal to those of other jobs. This simple method encodes each category (job title) independently from the others, and therefore fails to capture similarities between them. For instance, "Police Captain" is more similar to "Police Cadet" than "Property Manager II" and thus they are more likely to have similar salaries. Preserving this information when encoding categories is crucial to the performance of the machine learning model.

Dirty-cat is a Python package that provides tools to easily encode categorical data into numerical features while capturing similarities between categories: after encoding, similar categories are described by vectors with close coefficients, thus improving prediction. The package provides a user-friendly interface to easily transform a raw table with string features into a clean table that can be directly leveraged by machine learning models.

We will focus on methods developed in "Similarity encoding for learning with dirty categorical variables" and "Encoding high-cardinality string categorical variables" by Patricio Cerda et al., and implemented in the Gamma-Poisson encoder, the similarity encoder, the min-hash encoder, and the super-vectorizer.