A Benchmark and Taxonomy of Categorical Encoders

New. Comprehensive. Extendable.

A large share of datasets contain categorical features. For example, out of 665 datasets on the UC Irvine Machine Learning Repository [1], 42 are fully categorical and 366 are reported as mixed. However, distance-based ML models require features in a numerical format. Categorical encoders replace the categories in such features with real numbers.

A variety of categorical encoders exist, but there have been few attempts to compare them on many datasets, …

aia author benchmark categorical categorical-data data science datasets example features format however image machine machine learning mixed ml models numerical recraft taxonomy uc irvine