Commands and Operators

PREDICT and FOR EACH

A Predictive Query starts with a PREDICT clause that defines the value you want to predict. After specifying the prediction target in PREDICT, you must include a FOR EACH clause to specify the entity you want to make predictions for. Only a primary key column can be selected for this field.

Boolean Operators

You can use any number of boolean operators in your predictive queries. For example, LAST(LOAN.AMOUNT, 0, 30) = 2.

ASSUMING

ASSUMING allows you to investigate hypothetical scenarios and evaluate the impact of your actions or decisions. TheASSUMING keyword is followed by one or more future-looking temporal aggregations which will be assumed to be true during predictions.

For example, you may want to investigate how much a user will spend in the next 30 days if you give them more than two coupons or notifications in the next 7 days:

PREDICT SUM(TRANSACTIONS.PRICE, 0, 30)
FOR EACH CUSTOMERS.CUSTOMER_ID
ASSUMING COUNT(NOTIFICATIONS.*, 0, 7) > 2

📘

When generating training data, this acts as a forward-looking entity filter, and when we generate Batch Predictions, the model assumes that this condition would be met

CLASSIFY and RANK TOP K

When creating Predictive Queries with a LIST_DISTINCT aggregation target, you can use CLASSIFY and RANK TOP K to classify or retrieve only the top ranked values for a prediction, respectively.

CLASSIFY

For example, the following predictive query predicts customer purchases over the next 30 days, resulting in a separate binary classification per article ID:

PREDICT LIST_DISTINCT(TRANSACTIONS.ARTICLE_ID, 0, 30) CLASSIFY
FOR EACH CUSTOMERS.CUSTOMER_ID

If you don't specify a TOP K clause, we will generate classes for all available classes up to the limit.

RANK

In contrast, the following predictive query uses RANK TOP K at the end of the target definition (where K is the number of items of interest) to predict likely customer purchases—in this case ranking the top 12 products the customer is most likely to buy in the next 30 days:

PREDICT LIST_DISTINCT(TRANSACTIONS.ARTICLE_ID, 0, 30) RANK TOP 12
FOR EACH CUSTOMERS.CUSTOMER_ID

Using RANK on a foreign key is the only scenario where Kumo will allow adding new targets at batch prediction time—in this case, by adding new rows to the article table.