The ML.ENTITY_FEATURES_AT_TIME function

This document describes the ML.ENTITY_FEATURES_AT_TIME function, which lets you use multiple point-in-time cutoffs for multiple entities when retrieving features, because features can have time dependencies if they include time-sensitive data. To avoid data leakage, use point-in-time features when training models and running inference.

Use this function to retrieve features from multiple entities for multiple points in time. For example, you could retrieve features created at or before three different points in time for entity 1, and features created at or before yet another different point in time for entity 2. Use the ML.FEATURES_AT_TIME function to use the same point-in-time cutoff for all entities when retrieving features.

Syntax

ML.ENTITY_FEATURES_AT_TIME(
   { TABLE `PROJECT_ID.DATASET.FEATURE_TABLE_NAME` | (FEATURE_QUERY_STATEMENT) },
   { TABLE `PROJECT_ID.DATASET.ENTITY_TIME_TABLE_NAME` | (ENTITY_TIME_QUERY_STATEMENT) },
   [, NUM_ROWS => INT64][, IGNORE_FEATURE_NULLS => BOOL])

Arguments

ML.ENTITY_FEATURES_AT_TIME takes the following arguments:

• PROJECT_ID: the project that contains the resource.
• DATASET: the BigQuery dataset that contains the resource.

• FEATURE_TABLE_NAME: a STRING value that specifies the name of the BigQuery table that contains the feature data. The feature table must contain the following columns:

  • entity_id: a STRING column that contains the ID of the entity related to the features.
  • One or more feature columns.
  • feature_timestamp: a TIMESTAMP column that identifies when the feature data was last updated.

  The column names are case-insensitive. For example, you can use a column named Entity_ID instead of entity_id.

  The feature table must be in wide format, with one column for each feature.

• FEATURE_QUERY_STATEMENT: a STRING value that specifies a GoogleSQL query that returns the feature data. This query must return the same columns as feature_table. See GoogleSQL query syntax for the supported SQL syntax of the FEATURE_QUERY_STATEMENT clause.

• ENTITY_TIME_TABLE_NAME: a STRING value that specifies the name of the BigQuery table that maps entity IDs to feature lookup times. The entity time table must contain the following columns:

  • entity_id: a STRING column that contains the entity ID.
  • time: a TIMESTAMP column that identifies a point in time to use as a cutoff time when selecting features for the entity represented by the entity ID.

  The column names are case-insensitive. For example, you can use a column named Entity_ID instead of entity_id.

  The table identified by the ENTITY_TIME_TABLE_NAME value must be no larger than 100 MB.

• ENTITY_TIME_QUERY_STATEMENT: a STRING value that specifies a GoogleSQL query that returns the entity time data. This query must return the same columns as the ENTITY_TIME_TABLE_NAME argument. See GoogleSQL query syntax for the supported SQL syntax of the ENTITY_TIME_QUERY_STATEMENT clause.

• NUM_ROWS: an INT64 value that specifies the number of rows to return for each row in ENTITY_TIME_TABLE_NAME. Defaults to 1.

• IGNORE_FEATURE_NULLS: a BOOL value that indicates whether to replace a NULL value in a feature column with the feature column value from the row for the same entity that immediately precedes it in time. For example, for the following feature table and entity time table:

  Feature table

  +-----------+------+------+--------------------------+
  | entity_id | f1   | f2   | feature_timestamp        |
  +-----------+------+------+--------------------------+
  | '2'       | 5.0  | 8.0  | '2022-06-10 09:00:00+00' |
  +-----------+------+------+--------------------------+
  | '2'       | 2.0  | 4.0  | '2022-06-10 12:00:00+00' |
  +-----------+------+------+--------------------------+
  | '2'       | 7.0  | NULL | '2022-06-11 10:00:00+00' |
  +-----------+------+------+--------------------------+
  

  Entity time table

  +-----------+--------------------------+
  | entity_id | time                     |
  +-----------+--------------------------+
  | '2'       | '2022-06-11 10:00:00+00' |
  +-----------+--------------------------+
  

  Running this query:

  SELECT *
  FROM
    ML.ENTITY_FEATURES_AT_TIME(
      TABLE mydataset.feature_table,
      TABLE mydataset.entity_time_table,
      num_rows => 1,
      ignore_feature_nulls => TRUE);

  Results in the following output, where the f2 value from the row for entity ID 2 that is timestamped '2022-06-10 12:00:00+00' is substituted for the NULL value in the row timestamped '2022-06-11 10:00:00+00':

  +-----------+------+------+--------------------------+
  | entity_id | f1   | f2   | feature_timestamp        |
  +-----------+------+------+--------------------------+
  | '2'       | 7.0  | 4.0  | '2022-06-11 10:00:00+00' |
  +-----------+------+------+--------------------------+
  

  If there is no available replacement value, for example, where there is no earlier row for that entity ID, a NULL value is returned.

  Defaults to FALSE.

Output

ML.ENTITY_FEATURES_AT_TIME returns the input table rows that meet the point-in-time cutoff criteria, with the feature_timestamp column showing the timestamp from the time column of the entity time table.

Because you can specify multiple points in time from which to retrieve features for the same entity, it is possible to return duplicate rows, depending on the timestamps in the feature and entity time tables, and the num_rows value you specify. For example, if the only row in the feature table for entity ID 1 is timestamped 2022-06-11 10:00:00+00, and you have two rows for entity ID 1 in the entity time table that both have later timestamps, the function output has 2 rows with the same feature data for entity ID 1.

If either of the following conditions are true:

• No entity ids from the entity time table are found in the feature table.
• The rows in the feature table whose entity ids match those in the entity time table don't meet the point-in-time criteria.

Then the function doesn't return any output for that entity time table row.

Examples

Example 1

This example shows a how to retrain a model using only features that were created or updated before the timestamps identified in mydataset.entity_time_table:

CREATE OR REPLACE
  `mydataset.mymodel` OPTIONS (WARM_START = TRUE)
AS
SELECT * EXCEPT (feature_timestamp, entity_id)
FROM
  ML.ENTITY_FEATURES_AT_TIME(
    TABLE `mydataset.feature_table`,
    TABLE `mydataset.entity_time_table`,
    num_rows => 1,
    ignore_feature_nulls => TRUE);

Example 2

This example shows a how to get predictions from a model based on features that were created or updated before the timestamps identified in mydataset.entity_time_table:

SELECT
  *
FROM
  ML.PREDICT(
    MODEL `mydataset.mymodel`,
    (
      SELECT * EXCEPT (feature_timestamp, entity_id)
      FROM
        ML.ENTITY_FEATURES_AT_TIME(
          TABLE `mydataset.feature_table`,
          TABLE `mydataset.entity_time_table`,
          num_rows => 1,
          ignore_feature_nulls => TRUE)
    )
  );

Example 3

This is a contrived example that you can use to see the output of the function:

WITH
  feature_table AS (
    SELECT * FROM UNNEST(
      ARRAY<STRUCT<entity_id STRING, f_1 FLOAT64, f_2 FLOAT64, feature_timestamp TIMESTAMP>>[
        ('id1', 1.0, 1.0, TIMESTAMP '2022-06-10 12:00:00+00'),
        ('id2', 12.0, 24.0, TIMESTAMP '2022-06-11 12:00:00+00'),
        ('id1', 11.0, NULL, TIMESTAMP '2022-06-11 12:00:00+00'),
        ('id1', 6.0, 12.0, TIMESTAMP '2022-06-11 10:00:00+00'),
        ('id2', 2.0, 4.0, TIMESTAMP '2022-06-10 12:00:00+00'),
        ('id2', 7.0, NULL, TIMESTAMP '2022-06-11 10:00:00+00')])
  ),
  entity_time_table AS (
    SELECT * FROM UNNEST(
      ARRAY<STRUCT<entity_id STRING, time TIMESTAMP>>[
        ('id1', TIMESTAMP '2022-06-12 12:00:00+00'),
        ('id2', TIMESTAMP '2022-06-11 10:00:00+00'),
        ('id1', TIMESTAMP '2022-06-10 13:00:00+00')])
  )
SELECT *
FROM
  ML.ENTITY_FEATURES_AT_TIME(
    TABLE feature_table, TABLE entity_time_table, num_rows => 1, ignore_feature_nulls => TRUE);