How Categorical Data is Handled in Tabular Models

​

Overview

​

How It Works

1. Identifies categorical columns — any column set to a string/categorical type via the Data Types step is recognised as categorical.
2. One-hot encodes them — each unique category value becomes its own binary column (1 if that row matches the category, 0 otherwise). All unique categories present in your training data are encoded.
3. Drops one category per column — to avoid redundancy, the first category (alphabetically) is excluded. For example, if a column has values ["aluminium", "copper", "steel"], only two binary columns are created internally (copper and steel). When both are 0, the model knows the value is aluminium.
4. Scales your original numeric columns — your non-categorical input columns (e.g., temperature, pressure) are automatically centred and scaled so that features with different ranges contribute equally to the model. The binary columns created by one-hot encoding are not scaled.

​

How to Use Categorical Data

1. Import your data using the Tabular Loader.
2. Check column types in the Data Types step — ensure your categorical column is set to Categorical (string) type. The platform usually detects this automatically, but you can adjust it here if needed.
3. Add the categorical column as a model input — select it alongside your numeric columns when configuring the model.
4. Train the model — the platform handles the encoding internally.

​

Things to Be Aware Of

• All categories are used. Every unique value in the categorical column is encoded during training. If a column has many unique values (high cardinality), this increases the number of features the model sees internally.
• One category is always dropped. The first category alphabetically is excluded to avoid redundancy. This is particularly important for Linear and Polynomial Regression where including all categories can cause training issues. For tree-based models like Random Forest, this has minimal impact.
• New categories at prediction time. If a category appears during prediction that was not present in the training data, the platform handles this gracefully — the unknown category is treated as if none of the known categories match. Predictions can still be made without errors.
• Feature importance may not be available. When categorical columns are included, the number of features the model sees internally differs from the number of input columns you selected. As a result, per-column feature importance values may not be displayed for some model types.