Scholarlink

<p>Machine learning models often require numerical inputs, making the encoding of categorical features a critical step in the data preprocessing pipeline. A wide range of encoding methods, such as the commonly used one-hot encoding, are available, but they may not always be optimal due to increased dimensionality and a lack of sensitivity to the inherent relationships between categories. This paper presents a comprehensive evaluation of 26 categorical encoding techniques, benchmarked across 13 real-world datasets and 7 different machine learning algorithms. Our study categorizes these methods based on predictive task type, model performance, and computational efficiency, offering a taxonomy for selecting encoders. In addition, we illustrate how Safe AI metrics can be applied to encoding pipelines, showing that they provide complementary insights into model robustness and fairness. Finally, we provide a Python tool called <Emphasis FontCategory="NonProportional">EncodeHero</Emphasis>&#xa0;that enables researchers and practitioners to (1) extend the results by augmenting the benchmark with their own data and (2) choose the best encoding methodology based on their data and technical constraints.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Categorical variable encoding methods for tabular data: a benchmarking study

  • 期刊论文

  • 发表日期：

    2026年2月4日

查看全文
  • Federico Clerici,
  • Navid Nobani

摘要

Machine learning models often require numerical inputs, making the encoding of categorical features a critical step in the data preprocessing pipeline. A wide range of encoding methods, such as the commonly used one-hot encoding, are available, but they may not always be optimal due to increased dimensionality and a lack of sensitivity to the inherent relationships between categories. This paper presents a comprehensive evaluation of 26 categorical encoding techniques, benchmarked across 13 real-world datasets and 7 different machine learning algorithms. Our study categorizes these methods based on predictive task type, model performance, and computational efficiency, offering a taxonomy for selecting encoders. In addition, we illustrate how Safe AI metrics can be applied to encoding pipelines, showing that they provide complementary insights into model robustness and fairness. Finally, we provide a Python tool called EncodeHero that enables researchers and practitioners to (1) extend the results by augmenting the benchmark with their own data and (2) choose the best encoding methodology based on their data and technical constraints.