Conclusion

In this tutorial you trained a purchase propensity model on Signals attributes and scored live sessions with it. Along the way you:

• Defined, tested, and published a domain_sessionid attribute group and service, with the Python SDK or an AI assistant
• Built a labeled, point-in-time correct training dataset in Snowflake from the published definition
• Trained a gradient-boosted model, and evaluated it on held-out sessions
• Served the model behind a /predict endpoint that reads the same attributes live from Signals

The published attribute group defined the features for both training and serving, so the only feature code you wrote was the shared features.py.

Next steps

To improve the model and use it in production:

• Add features and compare the held-out metrics, for example user-level attributes keyed on domain_userid such as previous purchases, which the dataset builder joins to each anchor at the same point in time
• Try other dataset options, such as min_events to skip the first few events of a session, or max_negative_ratio to change the balance of the dataset, described in Create ML training datasets
• Review or customize the generated SQL with build_dataset_with_session_anchors(), if you want to run the dataset build on your own schedule
• Trigger actions without polling by using interventions, for example when a session adds to cart
• Follow the prospect scoring tutorial for a similar model on a marketing website