SAS Institute has unveiled Model Manager, a new application that promises to provide enterprises with greater control and standardisation over their internal processes for developing and rolling out predictive analysis and data mining applications. Implementing and managing the creation, management, deployment, maintenance and retirement of large numbers of models - something that SAS refers to as the 'analytic lifecycle' - is challenging in terms of time and resources.
This is especially the case for companies in financial services and life sciences industries that typically implement and maintain tens of thousands of models into the fabric of their operational systems.
"Leading organisations understand the requirement to treat the management of analytics as a lifecycle," said Keith Collins, chief technology officer at Cary, North Carolina-based SAS. He said the risk of 'model decay' is a serious inhibitor for large-scale deployment of predictive analytics in operational system. "Retaining poorly performing models in production status leads to inaccurate projections that are bad for business."
This decay, Collins explained, is often caused by the lack of a standardised iterative framework for model management that results in tedious, error-ridden model creation input as well as the incorrect interpretation of results.
Enter SAS Model Builder, a product that adds discipline and rigour to the process by introducing a best practices layer of checks and balances that continually verifies the accuracy and usefulness of analytic models in production.
Under the covers, Model Builder provides an integrated GUI environment for organising and tracking the cycle of model development, verification, testing, comparative performance benchmarking, publishing and sharing. Specifically it includes project-management templates, workflow, metadata gathering audit trail tools to provide guidance, documentation and collaboration around the lifecycle. The application is built on a core metadata repository that stores documentation about the model and scoring code information. This repository can also deploy the scoring engine from SAS Enterprise Data Integration Server or from any other components of the SAS Enterprise Intelligent Platform that produce analytical scoring engines.
A key feature of Model Manager is its metadata templates that are attached to every model. This allows users to input useful documentation about the history of the model - ie model metadata - for tracking and audit purposes. And because Model Manager is linked to SAS' Enterprise Miner data mining tool, visual icons can simply be dropped into the workbench to surface this metadata from the repository.
SAS has also licensed technology from Xythos Software to provide e-mail and Web alerting capabilities. There are also built in reporting and analysis tools that let users identify which models have best ROI and most accurate outputs. "Model Manager is not designed to build data mining models but to manage the models as more businesses look to implement predictive analytics into their decision making processes," explained Mary Crissey, analytics product marketing manager at SAS. "Building a predictive model is a complex process. But it is one that you need to get right. One error is scoring logic can potentially cost a company millions of dollars."
She said Model Manager also helps to bridge the gap between those who develop predictive and data mining models, typically statisticians and analysts, and how they are deployed in operational business environments.
Crissey added that managing the lifecycle of models has been "a very manual and time consuming process not to mention expensive as modeling skills do not come cheap. Companies are drowning in models and those who develop them are in hot demand. Yet they are spending a lot of their valuable time making sure that their models are being used, and used correctly, in various business groups."
Crissey explained that models often needed to traverse across different computing platform and sometimes needed to be recoded. "Many SAS modellers still prefer to work in our STAT tools and other legacy products to build predictive models. But these models need to be coded into different languages to be deployable operationally. Up to now this has been done in an ad hoc way and invariably errors creep into the translation."
Model Manager's ability to maintain metadata and comments about the model has been one of the valuable pieces that have been missing according to Crissey. "[Model Manager] is a useful communication and collaboration tool. It speaks both IT and business languages, bringing those two sets of users onto the same page."
Model Manager will ship in early November. SAS is also considering opening up Model Manager to non-SAS predictive and data mining models and leveraging open standards like PMML and UIMA. SAS also hinted the software would be updated to more easily import SPSS Clementine data mining models and provide better support for realtime deployment.
Despite focusing on high-end analytics, Crissey said that Model Manager is equally applicable to companies of all sizes. "Of course it will be most beneficial to any large enterprise that manages tens of thousands of models, like many of our financial services customers have to do. But companies that only have five to 10 models will also benefit as these models might be very complex to change."
Pricing for Model Manager starts at $75 000 upwards for a standalone solution, depending on platform and architecture. SAS is also offering Model Manager as a bundled component of larger enterprise-class Enterprise Miner suite deals.
SAS unveiled the product at the M2006 conference that ran in Las Vegas earlier this week. M2006 is billed as the world's largest data mining gathering in the world.