Snowflake Unlocks Multiparty ML in Data Clean Rooms

Snowflake is moving beyond basic SQL queries within its Data Clean Rooms. The company announced today that ML Jobs, its feature for running sophisticated machine learning workloads, is now generally available. This advancement allows data scientists to bring their familiar Python ML stacks, complete with distributed training, hyperparameter optimization, and GPU acceleration, directly into multiparty data collaborations.

Previously, data clean rooms were often bottlenecked by limitations to SQL or single-node Python, hindering enterprise-scale ML. ML Jobs aims to transform these environments from mere compliance tools into active hubs for model building. Organizations can now train models on combined data from multiple parties without exposing raw records, automating complex pipelines.

Consider advertising: an advertiser might need publisher ad log data, identity provider signals, and retail transaction data to build robust audience and measurement models. Each party has valid concerns about data privacy and intellectual property. ML Jobs addresses this by ensuring data providers govern their information for explicitly approved workloads, while the advertiser's proprietary model logic remains within the secure collaboration boundary.

This capability is foundational for future AI advancements, such as training AI agents that rely on signals distributed across various organizations. The collaborative machine learning approach, powered by Snowflake's infrastructure, is set to redefine how AI operates across company lines.

Multiparty Model Training and Scoring

Machine learning models trained on isolated data silos offer a limited perspective. By combining distinct signals, like purchase history from retailers, transaction patterns from financial services, and engagement data from brands, models become significantly more predictive. ML Jobs makes it practical to run unified training pipelines across these varied feature sets.

Propensity scoring exemplifies this. A model trained solely on publisher behavioral signals is less effective than one that incorporates advertiser first-party conversion history and data provider demographic enrichment. The resulting scores are more accurate because the model has a comprehensive view of conversion drivers.

Simplified Development and Scalable Deployment

For data scientists, the development experience mirrors their existing workflows. They can write standard Python code, utilize preferred IDEs, and specify requirements via a simple YAML configuration. There's no need for complex Docker image builds or manual infrastructure provisioning.

Scaling compute resources, including to multiple nodes or GPUs, becomes a parameter adjustment rather than a fundamental architectural change. Workloads are designed for production, allowing for scheduled runs, event triggers, or orchestration via standard tools.

Iteration occurs in familiar development environments before seamless deployment into the clean room, ensuring a smooth operationalization process. Audit trails and queryable activity history provide transparency and debuggability.

Key Use Cases Emerge

Incrementality measurement, crucial for understanding advertising's true sales lift, historically required a neutral third party or custom infrastructure. With ML Jobs, brands and retailers can now run uplift models directly within the collaboration, keeping impression logs and transaction data in their respective accounts.

Retail media attribution at scale is another key area. Transaction data, highly valuable for attribution, has been difficult for agencies to access. ML Jobs enables attribution models to run where the data resides, with only model outputs shared downstream.

As third-party cookies fade, identity crosswalks are vital for maintaining match rates. ML Jobs facilitates probabilistic identity resolution by training models on combined advertiser CRM data and identity provider graphs without data leaving accounts. This approach recovers significant match rate lift and adapts to evolving ID coverage.

The platform also supports advanced applications like campaign optimization agents. Unlike static lookalike models, these agents reason over combined signals from multiple parties to recommend targeting strategies, budgets, and bid levels. This level of collaborative machine learning requires the trust and governance provided by clean room environments.

Snowflake's ML Jobs differentiate themselves by supporting end-to-end Python ML workflows optimized for automated production pipelines, contrasting with platforms focused on specific use cases or shared notebooks. The system's hash-based approval and Cross-Cloud Auto-Fulfillment capabilities further streamline operations across different cloud environments.

ML Jobs in Data Clean Rooms is available now for all Snowflake accounts with the Data Clean Rooms environment installed. Example workflows for lookalike audience modeling and multiparty incrementality measurement are provided to help users get started.