I wrote about FICO Model Central, FICO’s model management product, back in February of 2012. A new release, 3.0, is shipping with a new feature called Model Deployment Accelerator for rapid deployment.
To recap, FICO Model Central automates analytic model compliance tasks, provides alerts and visualization to help you identify needed model updates, supports getting multiple vendors’ models into production, eases embedding models into decision services and integrates optimization and simulation. It comes in incrementally adoptable Foundation, Professional Development and Advanced Decisioining tiers. Model Central can manage models whether or not the new features of Model Central are used to deploy the model.
New in 3.0 is the Model Deployment Accelerator, a capability for rapid deployment of models expressed as SAS code, as PMML or loaded from FICO Model Builder, to scoring services. Model Deployment Accelerator allows models to be quickly deployed as a service, eliminating the time and potential errors involved in reprogramming the model for production. For each model the user can see every version of the scoring services generated for that model. Configuring a new service requires a model definition file, a Base SAS file for instance, and a pair of audit files (input data and expected outputs for that input data). Uploading the SAS script to Model Central converts it to Java which is then encapsulated into a callable service. These scoring services can then be integrated into a web-managed rules-based Decision Service. SAS files can include pre-processing, variable calculation, transformations, segmentation logic and other script elements as well as the scoring formula and reason code calculations.
Once the Java scoring service is generated, the uploaded audit files (or a sample from those files) can be run through the code to confirm that the results are what the user expected. An audit report is generated that highlights any differences between expected and actual results that are outside a tolerance specified by the user. The report can display just the rows with problems, helpful when there are lots of test rows. If there are problems then a corrected scoring program or audit file can be uploaded, and the test repeated.
Once a user is satisfied that the service is scoring correctly, they can deploy the model as either a standalone Java service or as a Java packet suitable for use inside FICO Blaze Advisor (FICO’s business rules management system). Once this is generated the version of the scoring service is locked so it cannot be changed (keeping verification reports, code etc available for future auditing and reference). The generated JAR file, as well implementation documentation and some supporting Java libraries , are packaged up into a zip file ready for deployment. Once the scoring service is actually deployed, or subsequently if it is retired, Model Central also allows the user to manage additional status levels, such as “In production”.
You can get more information on FICO Model Central here and FICO is one of the vendors in our Decision Management Systems Platform Technology Report.