Question

A data science team at a retail company wants to train a tabular classification model in Vertex AI to predict customer churn. Their source data lives in a BigQuery table containing 3 years of transactions, including a purchase_timestamp column. They want Vertex AI to manage the data splits and preprocessing metadata, and they are concerned that a purely random split could cause the model to learn from future events when predicting past behavior. Which approach should they use when creating the Vertex AI managed dataset and training job?

Accepted Answer

Create a Vertex AI managed tabular dataset from the BigQuery table and configure a chronological (time-based) split using the purchase_timestamp column so earlier data trains the model and later data validates and tests it.

Answer

Export the BigQuery table to CSV in Cloud Storage and let Vertex AI apply the default 80/10/10 random split during training.

Answer

Create the managed dataset and use a manual split, assigning rows to TRAIN, VALIDATION, and TEST uniformly at random with a hashed customer_id column.

Answer

Shuffle the BigQuery table with a random ORDER BY, then create the managed dataset so Vertex AI performs a stratified split on the churn label.