Automatic feature preprocessing

BigQuery ML performs automatic preprocessing during training by using the CREATE MODEL statement. Automatic preprocessing consists of missing value imputation and feature transformations.

For information about feature preprocessing support in BigQuery ML, see Feature preprocessing overview.

For information about the supported SQL statements and functions for each model type, see End-to-end user journey for each model.

Missing data imputation

In statistics, imputation is used to replace missing data with substituted values. When you train a model in BigQuery ML, NULL values are treated as missing data. When you predict outcomes in BigQuery ML, missing values can occur when BigQuery ML encounters a NULL value or a previously unseen value. BigQuery ML handles missing data differently, based on the type of data in the column.

Feature transformations

By default, BigQuery ML transforms input features as follows:

TIMESTAMP feature transformation

The following table shows the components extracted from TIMESTAMP columns and the corresponding transformation method.

Category feature encoding

For features that are one-hot encoded, you can specify a different default encoding method by using the model option CATEGORY_ENCODING_METHOD. For generalized linear models (GLM) models, you can set CATEGORY_ENCODING_METHOD to one of the following values:

• ONE_HOT_ENCODING
• DUMMY_ENCODING
• LABEL_ENCODING
• TARGET_ENCODING

One-hot encoding

One-hot encoding maps each category that a feature has to its own binary feature, where 0 represents the absence of the feature and 1 represents the presence (known as a dummy variable). This mapping creates N new feature columns, where N is the number of unique categories for the feature across the training table.

For example, suppose your training table has a feature column that's called fruit with the categories Apple, Banana, and Cranberry, such as the following:

In this case, the CATEGORY_ENCODING_METHOD='ONE_HOT_ENCODING' option transforms the table to the following internal representation:

One-hot encoding is supported by linear and logistic regression and boosted tree models.

Dummy encoding

Dummy encoding is similar to one-hot encoding, where a categorical feature is transformed into a set of placeholder variables. Dummy encoding uses N-1 placeholder variables instead of N placeholder variables to represent N categories for a feature. For example, if you set CATEGORY_ENCODING_METHOD to 'DUMMY_ENCODING' for the same fruit feature column shown in the preceding one-hot encoding example, then the table is transformed to the following internal representation:

The category with the most occurrences in the training dataset is dropped. When multiple categories have the most occurrences, a random category within that set is dropped.

The final set of weights from ML.WEIGHTS still includes the dropped category, but its weight is always 0.0. For ML.ADVANCED_WEIGHTS, the standard error and p-value for the dropped variable is NaN.

If warm_start is used on a model that was initially trained with 'DUMMY_ENCODING', the same placeholder variable is dropped from the first training run. Models cannot change encoding methods between training runs.

Dummy encoding is supported by linear and logistic regression models.

Label encoding

Label encoding transforms the value of a categorical feature to an INT64 value in [0, <number of categories>].

For example, if you had a book dataset like the following:

The label encoded values might look similar to the following:

The encoding vocabulary is sorted alphabetically. NULL values and categories that aren't in the vocabulary are encoded to 0.

Label encoding is supported by boosted tree models.

Target encoding

Target encoding replaces the categorical feature value with the probability of the target for classification models, or with the expected value of the target for regression models.

Features that have been target encoded might look similar to the following example:

# Classification model
+------------------------+----------------------+
| original value         | target encoded value |
+------------------------+----------------------+
| (category_1, target_1) |     0.5              |
| (category_1, target_2) |     0.5              |
| (category_2, target_1) |     0.0              |
+------------------------+----------------------+

# Regression model
+------------------------+----------------------+
| original value         | target encoded value |
+------------------------+----------------------+
| (category_1, 2)        |     2.5              |
| (category_1, 3)        |     2.5              |
| (category_2, 1)        |     1.5              |
| (category_2, 2)        |     1.5              |
+------------------------+----------------------+

Target encoding is supported by boosted tree models.