Announcing RudderStack Predictions: Automate churn and conversion scores in your warehouse

ℹ️ Predictions is now the Propensity Scores Data App. Read more in the announcement blog.

Machine learning is now central to personalized customer experiences. It’s quickly becoming table stakes if you want to build and maintain a competitive advantage in today’s market. However, while the rapid injection of AI into our daily lives accelerated the push for every company to leverage ML to its full extent, most companies are struggling to deliver.

If your efforts to drive business outcomes with ML are stalled because of the challenges around creating and deploying ML models, you’re not alone. KDnuggets reported that 80% of ML projects fail before deployment.

We’re introducing RudderStack Predictions to address this problem head-on. With Predictions, you can now quickly train and automatically deploy ML models for churn and conversion scores without complex, expensive MLOps infrastructure. It seamlessly integrates into your existing workflows so you can deliver ML outcomes with velocity, and it’s now available to RudderStack Enterprise customers.

Drive business outcomes fast with automated ML

We created RudderStack Profiles to streamline identity resolution and customer 360 creation. Now, we’re doing the same for ML models. Predictions significantly reduces the time required to train and deploy ML models for churn and conversion. It enables you to target and personalize based on anticipated customer behavior so your business teams can take proactive action to reduce churn and increase conversion.

How RudderStack Predictions works

Predictions operates within the same workflow as Profiles, allowing non-data scientists to deploy ML models using their existing data engineering workflow. It leverages your data platform's existing ML compute infrastructure (such as Snowflake's Snowpark ML platform) and pushes results directly into your warehouse, making ML data immediately accessible through our Reverse ETL pipeline or Activation API.

Our warehouse-native approach ensures Predictions is fully auditable, so you can thoroughly audit models and fit metrics, runs, and outputs.

Using RudderStack data sources, Predictions automatically produces churn and conversion scores, seamlessly integrating them into your customer 360 table. But Predictions isn’t limited to RudderStack data sources. It can use any data you have in your warehouse. Plus, if you have advanced use cases, you can migrate to a version controlled, code-based workflow to create custom predictive features using any data in your warehouse.

YAML
Copy
predict:
  inputs: 
    - packages/base_features/models/rudder_user_base_features
  config:
    data: 
      package_name: feature_table
      label_column: large_purchase_last_90 
      label_value: 1 
      prediction_horizon_days: 90
      features_profiles_model:  'rudder_user_base_features'
      eligible_users: 'large_purchase_last_90 is not null'
    preprocessing: 
      ignore_features: [first_name, last_name, state]
    outputs:
      column_names:
        percentile: percentile_large_purchase_90
        score: likelihood_large_purchase_90

Getting Started

Predictions is now available for RudderStack Enterprise customers and prospects interested in testing the product.

We’ve partnered with Snowflake to build a detailed Quickstart Guide, complete with a sample data set and an explanation of how Predictions automatically leverages the power of Snowpark for ML. Check out the guide and reach out to our team to enable Predictions on your account or get a personalized demo.

You can also check out the docs for more details and dig into our Shopify churn model project on GitHub to see how Predictions builds a churn score from Shopify features.

This is just the beginning for Predictions. We’ll continually add new models for more use cases, so stay tuned. We’re even working on enabling you to migrate your existing ML models into Predictions for scoring and activation.