As new fraud trends continue to emerge in the post-pandemic, digital world, we are constantly evaluating and updating our machine learning models to ensure they provide the most accurate and performant data for customers.
Releasing new model versions to customers can present a few challenges: balancing frequency of updates with customer impact, timing of releases, and more. Furthermore, many of our customers use our model-powered attributes in high-volume, top-of-waterfall integrations to help assess risk in real-time. When these attributes change, there can be significant downstream effects if these updates are not properly planned for and if the necessary adjustments are not made.
With the feedback of our customers, we have developed a new model release approach that kicks off in August to help our customers plan for, and migrate to, the latest versions of Ekata’s machine learning model-powered attributes. This new, holistic approach to how we release model updates will balance the value of up-to-date models for our customers, while minimizing the cost that ingestion of these updates can occur. All aspects of a release will be addressed in this new approach – timing, communications, and the overall migration experience.
Here’s how the new process will work:
- Models will be released in February and August of each year
- Following the release, there will be a transition window of five months, between the current and newest models available during that time frame
- Customers will have full control over which model is being used in production, and can choose when to migrate during the five-month transition window to do so
- Customers can also choose a default behavior for when new models are released so the newest versions are immediately implemented when available
- Each release will be accompanied with data to help assess impact to production systems and help you make the appropriate adjustments
Our new model release approach will provide greater transparency around current versions and our newest model updates so customers can better understand how their business will be impacted. An update to Ekata’s models may cause the volumes of transactions in a given score range to change as a result of the model’s improved performance delineating risky and non-risky identities. Moving forward, we will be providing data from customer’s live transaction data with both the current and new model score distributions so these shifts can be identified and any changes required to accommodate can be calculated. (Example below.)
Admins can now manage models (for single and multiple API keys) in our self-service tool and see which:
- Model versions are currently available
- Default behavior has been set when new models are released
- Features are impacted by the latest model update. This can affect the following features in each endpoint
Our team here at Ekata is committed to delivering the best customer experience possible when releasing new models into production. Despite the challenges that can arise for customers when new models are released, this new approach to model retraining will provide predictability, consistency, and flexibility for customers when transitioning to new versions. To learn more, check out our model release approach FAQs.