SimilarityEncoder turns categorical strings into similarity-based features. Instead of treating every spelling as unrelated, it uses character or n-gram similarity to group high-cardinality categories and soften noisy values such as typos. section 8.4
A practical workflow is:
- Identify high-cardinality string columns before encoding them. section 8.4
- Use SimilarityEncoder for similarity-based one-hot-style features, while treating low-cardinality and continuous columns separately. section 8.4
- Place the encoder in a modeling pipeline and compare it with alternatives such as MinHashEncoder and GapEncoder. section 8.4
Employee salary prediction
In an employee-salary regression dataset, employee_position_title contains 385 categories. SimilarityEncoder is used with XGBoost alongside other dirty_cat encoders; the reported R² is 0.89956. section 8.4
