SPSS Inc. likes to say they focus on helping customers capture all the information they need, predict outcomes and then, using their Decision Management products, act on these insights by embedding analytic results into business processes. Within this family, the PASW Decision Management tools add actions, business rules, to analytics to enable action to be taken in a business process. The products available today handle such things as improving customer intimacy in customer retention, boosting product leadership in product placement and operational excellence in fraud detection.
There are typically many deployment challenges that SPSS meets by putting business users in control of decisions, integrating recommendations and automating deployment of analytics. Optimized decisions are made by the products based on analytics, business rules and a set of options to find the best one – business goals used to arbitrate the optimal decision.
The 3 decision management products are Event Builder for outbound marketing campaign optimization, Interaction Builder for inbound interaction decisions and Risk Control Builder for the claims handling process. Each has an interface for business users to define the parameters and tradeoffs and structure of decisions along with constraints and rules for the decision. They offer very much a classic decision management proposition. Each is focused on their specific business users and support simulation and monitoring to test multiple scenarios. All integrate with other systems – providing decisions, Decision Services, to other process-centric applications. In a claims process, for instance, you get first notification of loss and can then use Rick Control Builder to determine whether you should put the claim on the fast track for payment, request additional information or investigate the claim for fraud. As the process continues this can be repeated after more information is gathered or an in depth investigation is completed.
The products are focused on real-time or decision making, though they support batch decisions, and include the actual context – the interaction – as part of calling the engine. Rules for eligibility of action can be defined and additional rules to prioritize actions (e.g. by margin or risk) can be specified. Predictive models – likelihood of acceptance or of fraud – are imported from PASW Modeler and all this is combined to produce the best action. The applications are Java-based and designed to fit on anyone’s platform and use existing data sources, both structured and unstructured. They use the same data platform as the PASW Modeler, for instance, allowing them to share data definitions and they are intended for a business user to manage – no programming.
As an example, the campaigns product allows multiple campaigns to be defined and uses predictions about response, profitability, etc. to determine who gets assigned to which campaign when. This avoids over saturating good customers and makes sure customers get maximized results from minimum hits. For example, one customer began with 3 broad-based campaigns a year and looked to the Decision Management application to improve their results. First they used Predictive Analytics to target correctly and reduced the cost of their campaigns by 35%. Then they moved from 3 big campaigns to many smaller campaigns and used more analytics to target these to appropriate customers and doubled their response rate. This created overlapping campaigns hitting the same customer and so over saturated some customers. They used the optimal decision approach to optimize which offer went to whom and this led to 29% increase in profit.
The Risk Control Builder product handles all aspects of the claims process from a decisioning perspective. Different risk areas are defined at the granularity the user prefers – auto risk areas specific to regions for instance or subrogation risk. Business rules are defined and then fire to add risk points to each claim based on these different risk areas. Each rule that fires can add its reason text to the results to help explain the outcome. Models are deployed automatically and seamlessly. The SPSS layer handles transformations, missing data, etc. and is available both to the PASW Modeler and to the deployment management. Each analytically-derived rule influences the optimized risk score. Finally a user can define a tradeoff matrix between basic risk levels (from rules) and optimized risk levels (from models) and this determines the decision that will be returned for a claim.
Rules are pretty easy to create based on the available data. The product shows user friendly data element definitions including derived data from the SPSS layer and allows a user to pick attributes and then operators and known values from lists constrained by that attribute choice. For each rule conditions are defined along with the risk score outcome. Adding to the risk score is the only allowed action for a rule but otherwise these are pretty standard rule definitions. The product supports testing to allow test scenarios to be run and the results examined. It has a nice simulation facility that allows a whole set of historical data to be processed to see how the distribution would change from the current configuration and this can be compared to the expected/desired distribution. The user can change values in rules and see quickly what the effect would be. Versioning and an audit trail is available for all assets in the PASW platform, rules as well as models, and the platform allows a user to pull the assets for a particular time and date and use them in a scenario e.g. to support compliance.
While the rules in these products are not as broad-based as those in a typical business rules management solution, the seamless deployment of models as well as the focus on decisions not just predictions were very welcome. While I still think many decisions call for a true balance between business rules and analytics, this analytic-heavy deployment has a lot to offer many SPSS modeling customers.
