A little while ago I got to talk to Mike Zeller and his team at Zementis about their decision management platform – ADAPA, what they call a predictive analytics decision engine. ADAPA stands for Adaptive Decision And Predictive Analytics.
The folks at Zementis started trying to solve a fairly common problem in organizations adopting data mining and predictive analytic techniques – they have models that would be useful if only they could get them into their production systems. They therefore focus on an easy deployment approach for models. The execution platform supports both real-time and on demand batch scoring, has features to help you manage and check up on deployed models and uses SOA-based IT integration – open standards. They don’t have a model development environment- people use one of the established platforms like SAS or SPSS. The key value proposition for ADAPA seems to be a lower TCO as it can run on an open source stack, indeed it is very agnostic to the IT stack being used. The engine is a plug-and-play Java service ready for SOA.
The execution engine has a web console for management and monitoring of deployed models. The console has some nice reporting built in and some simple but fairly elegant rule and model management capabilities. The engine supports Web Service calls for bulk or single case execution. It accesses a repository and an audit database using JDBC, executes business rules and predictive models and returns results to calling applications. To bring in models, ADAPA imports PMML definitions and allows use of JSR 73 Java Data Mining APIs. Their standards support continues in that rules can bring in XSD data, reports can use XML, the whole thing can be executed through the Java Rule Engine API JSR 94. While some of these standards have limited penetration it is improving and Zementis is doing what it can to help. It also leverages open source – the rules engine is based on JBoss DROOLS (of which more here on the Drools blog) and the reporting infrastructure is based on Jasper.
From a modeling point of view they support Neural Net, Support Vector Machine and Regression models. They get around one of the complaints I have heard from folks using PMML – its lack of focus on data transformation – by using the PMML options for pre- and post-processing. This allows them to import all the data transformation logic.
They have extended Drools with Excel-based table-driven authoring. While they have customers using ADAPA purely for rules, they don’t have a general purpose rule management user interface. They offer a straightforward implementation for business logic so that you can combine models and segments, and design treatment approaches around the predictions of your models.
In addition, model execution is implicit (it only happens when a score result is required) and you can chain rulesets. Their Excel interface supports a lot of condition columns and has something of a Corticon look and feel. Once the rules are loaded up they have a nice little validation framework and the engine handles reason codes.
It’s a nice platform showing some nice rule and model integration, good standards support and some intelligent use of open source. They have posted a couple of demos recently at http://www.zementis.com/demo
- ADAPA Predictive Analytics Edition Demo: a demo of our predictive scoring engine
- PMML Converter: Converts PMML files into the latest version (3.2) of the PMML standard
These demos are also available as an iGoogle Gadget which shows how (potentially) simple it can be to leverage predictive analytics!