Chordiant Decision Management Update

I got a chance to sit down with Rob Walker last week for an update on Chordiant Decision Manager. Rob covered some of the new features in Chordiant Decision Management 6.2 as well as some general background that has not appeared in any of my posts before (check out First Look – Chordiant Decision Management for a basic overview and First Look – Chordiant’s Visual Business Director – VBD is generally available in this release).

I have not really discussed Predictive Analytics Director before. This takes business users through a wizard that builds models in a traditional offline approach – using historical data to build a model at a point in time. It is aimed at business users who are trying to predict customer behavior and develop scoring models. A highly automated, business focused tool designed to build models that predict customer behavior it generates a lot of additional information wrapped around the model to help users of the model. For instance it allows you to embed the classifier and business metrics not just the score. This means you don’t have to write rules that say “score > x” – you can say “risk is AAA” or “risk of churn is moderate” and these execute based on score. These traditional models are best used for risk and churn but for propensity models they recommend adaptive analytics.

Adaptive models change to in real-time as customers respond. These are built using variations on the naïve Bayes algorithm entirely from scratch in a dynamic way. These use sophisticated, real-time auto-binning and auto-grouping capabilities, as well as a very scalable architecture that lets you have hundreds of these models concurrently learn and predict in real-time. You can also run several time windows concurrently. You might favor, for instance, a slow learning model over a fast moving model when circumstances change because the slower model is more stable. Users define the proposition (decision) for which a propensity is required, optionally the number of classes or buckets you want in the results, the window of memory (transaction count) over which the model should be built and the inputs you want to consider. You can then classify the scores for the number of buckets you specified and link rules to the buckets to specify actions and so encapsulate the actions taken as a result.

Besides the two built-in model types, CDM also allows you to import PMML from other modeling environments. Once imported to Strategy Director, the core modeling environment, they can be used just like any other model from Predictive Analytics Director but lack some of the metadata and associated information created by the wizards. Strategy Director is often used by customers to arbitrate between the various models (adaptive models, traditional models and classification models) using rules. For instance, some users have a switch rule to see if the adaptive model is reliable enough (typically after a few weeks) and use a static model until it is.

Strategy Director allows you to use adaptive models, imported models and general models as well as various other decision components. The decision components include input components (fields, virtual fields and expressions), classification components (predictive model, adaptive model, classification rule (decision tree), matrix rules (decision table) or scorecard), modification components (mapping rules and exceptions), strategy components and selection components (prioritization, switch and champion/challenger rules).