Categorical encoding

Most machine learning algorithms require the variables to be numeric. Because a categorical value does not have a clearly measurable relationship with the other values in that same column, it must first get converted into a numeric representation to be measured by math. AutoML uses categorical encoding to transform categorical values in feature columns into numeric values that machine learning algorithms can understand.

AutoML uses two encoding methods: impact encoding and one-hot encoding. The method used on a particular feature depends on the dataset size and the number of unique categorical values.

• For datasets with 100 or fewer columns:

  • Categorical features with 13 or fewer unique values are one-hot encoded.

  • Categorical features with more than 13 unique values are impact encoded.

• For datasets with more than 100 columns, all categorical columns are impact encoded.

You can see which features in your dataset are being processed using categorical encoding by consulting the schema view when configuring your ML experiment. For more information, see Configuring experiments.

How does categorical encoding work

A common technique for giving mathematical representation to a category is one-hot encoding. One-hot encoding pivots the categorical column into n number of columns, where n is equal to the number of unique values in the column. The number 1 is assigned to the appropriate column for each row and 0 to the other columns that were generated for the category. Categorical encoding allows each unique variable to be evaluated independently of the others, unlike a numerical value that is evaluated in relative terms to the other values in the column.

The example in the table shows how the categorical column MarketingSource has been one-hot encoded. The result is four new columns—one for each unique marketing source. On the first row, Person_1 has marketing source "Facebook". This is represented by 1 in the new Facebook column, and 0 in the other columns.