- Mining enhancements like data weighting, in-database analytics, new statistics
- Support for more complex IT environments
Angoss has long supported the development of strategies or decision trees – models that define the relevant customer segments for a population and then assign treatments to those segments. The new optimization wizard allows miners to specify a linear objective function and constraints (risk, profit, bad debt limits etc) and then assigns optimized treatments to selected nodes in a strategy or decision tree. This function is in addition to a more rules-based approach where the rules for assigning treatments to customer segments are specified manually. Users previously exported the definition of a tree, ran it through an external optimizer and then put it back. This design-time optimization is now supported completely within Angoss.
Version 7.0 also makes data weighting available across the board – in all models, in creation of training data sets, data profiling, building trees etc. Users can use a wizard to create a weighting function or import a weight field with their data. This is a big deal for risk customers who often want to weight certain data more heavily. For instance, with the recent changes in the market, many want to weight data from recent periods differently from older historical data.
There is a growing demand for in-database analytics from the Angoss customer base. In addition some customers don’t like moving data from one location to another to create models as this creates data silos. Angoss 7.0 offers a new connection so that data can be sourced directly from databases. In addition, the model creation routines can be run in-database. A fairly generic engine has been developed and certified with Netezza and SQL Server already. They have kept this engine fairly generic to make it portable while recognizing that this means they can’t necessarily take full advantage of different platforms.
On the usability front they have redesigned the dataset partitioning wizard to give more information and better visualization of the partitions, added undo in various modeling tasks, added PDF generation for reporting and improved decision tree printing with scaling and page break management. In addition new dataset and model analysis statistics are available and the modeling server is now supported on Linux, Vista, Windows Server, XP, AIX and Solaris.
Finally 7.0 has added more model export formats. PMML, SAS code, and generic XML are now supported for all the models they support (decision trees, strategy trees, logistic and linear regression, MLN, Cluster and Scorecard). SPSS code, Java, SQL, text for reporting are also offered for decision trees and increasingly for strategy trees and scorecards. They are working to ensure that most model types can be exported in to a wider range of deployment languages. They see a particular growth in demand for PMML, as do I, and have more and more users adopting PMML to move models into production environments such as Business Rules Management Systems.
All in all some nice new features. I was particularly glad to see more support for PMML, making production deployment easier, and the direct database connection as these reduce the impedance between modeling and operational decisioning.