I received an update from IBM recently on IBM SPSS Decision Management 6 and IBM SPSS Modeler 14. IBM sees organizations with many different data sources (inside and outside) that are moving from a traditional approach focused on reporting to one that involves “predict and act” – real-time, fact-driven decision making at the point of impact.
IBM, like me, uses a view of decisions ranging from Strategic Direction, Managerial Control decisions (a better name for tactical decisions) and decisions about Operational Execution. As I have said before automation/decision management is focused on the Operational Execution decisions while decision support or offline focuses on the Strategic Direction decisions (with Managerial Control decisions getting a blend). In this context, IBM wants to move from delivering advanced analytics in reports and dashboards towards enabling the use of analytics to achieve a business end – a 1:1 relationship with the customer helping better deliver what customers might want at a specific time, through a specific channel.
Obviously, integrations with other IBM technology will be essential to the overall effectiveness of this vision. Such integration is obviously already possible to some degree using linkages through a SOA architecture as SOA support is common to all the relevant IBM products (WebSphere Process Server, WebSphere Business Events, Lombardi, ILOG Rules etc). According to IBM these integrations could become more productized over time.
IBM SPSS Decision Management then combines predictive analytics, business rules and optimization to automate high volume decisions made every day – those related to Operational Execution. Examples include deciding which cars to search at the border, which claims to investigate or which customers to save. This product embeds analytics deep within an organization’s DNA, pushing analytics right out to the front line. IBM has a repeatable process for this that they believe covers all such decisions:
- Connect to data
- Define global selections – ability to include / exclude certain data from the analysis.
For example, in the insurance industry, you may exclude “acts of God.” - Define desired outcomes – the options from which the decision should select
- Define operational decisions with rules and predictive models
- Optimize outcomes, simulate outcomes
- Deploy into operational systems
- Report
Of course, the business owner needs to be able to participate through appropriate user interfaces and needs to know which decisions they are thinking about before starting. Also, the business user has the ability to define / update the decisions that are delivered into the operational environment.
IBM SPSS Decision Management is available in three forms initially. It comes as a software platform for managing “content” that defines the decisions and two sets of pre-configured packages (software and content) for Claims and for Customer Interactions. Once the decision-specific content is created (by a customer or as part of a package) the software interface adapts based on the content (rules and models) – there is no problem-specific code. The steps are configurable (new ones can be added, irrelevant ones excluded), data can be pre-defined, rules can be included etc.
The rules interface for Decision Management is fairly simple (it is not as sophisticated as in a BRMS like IBM ILOG, but it can be integrated with these systems) and allows business people to manage the rules in it. These rules can reference a predictive analytic model (if model.score > X and … THEN …) so that rules and models can be used to make a decision in combination. It is also possible to define a set of rules and a set of models each of which select an action and then build a trade-off matrix between them.
The product also supports decision simulation and business users can change the rules’ parameters/thresholds before running the simulation to see how the distribution of selected actions changes. They can use test data or last month’s data so that they can do impact analysis – what would the impact have been last month if I did this instead of that?
The platform supports randomized offers for champion/challenger or A/B testing too. Business users can simulate multiple times and find the best approach by comparing results. It’s not the most sophisticated decision simulation environment, but it’s a good beginning. In addition to configuring and testing the decision, a business user can use IBM SPSS Modeler Advantage to add new models of which more in a minute.
For data mining and text analytics, IBM SPSS Modeler 14 is focused on automation, enterprise support, new analytics and supporting different kinds of users. Modeler has a typical data mining/predictive analytics interface aimed at modelers – nice graphical view, but still potentially very complex given the number of steps involved in cleaning data, building the model, comparing and integrating outputs etc.
New auto nodes handle common activities like data preparation (turning absolute dates into time since someone became a customer for instance), allow someone to say what they want to predict and allow the tool to select the modeling techniques appropriate. This dramatically reduces the complexity of the streams.
With Modeler Advantage the interface is very simple so that business users can build a model without the skills of a typical Modeler user. Modeler Advantage allows business users to select data, build models and apply models in the decision management framework without any great level of analyst skill. Rather than being standalone, Modeler Advantage generates the model creation stream behind the scenes. This shared underlying model allows modelers and business people to work together: the modeler adds expertise to those streams generated by business users, for instance, or integrates standard elements already in use within the company. This is a nice sign that the analytic and business communities are starting to figure out how to share analytic model creation – just as IT and business communities have had to figure out how to share business rules/business logic creation.
Modeler 14 also improves the integration of text analytics, something SPSS has focused on for a while, with new pre-configured libraries defining everything from finance words to emoticons. Modeler 14 continues to support IBM, Microsoft and Oracle in-database algorithms. It also supports SQL Pushback, pushing data preparation tasks back to the database as SQL, and allows very large datasets to be used with models being built with incremental refresh.
The end result of Modeler 14 and Decision Management 6 is a common set of underlying modeling algorithms and capabilities with interfaces for modelers, less experienced modelers and even business users. And once the decisions are described they can be deployed using Decision Management. For instance, a business user builds a model, mapped to an underlying modeler stream, which can then be used in champion/challenger testing, etc. using the decision management platform. Once the results justify it, the model can be made available for production deployment and the results of using the model collected and fed back into the process to close the loop.
Update: You can read the press release here and get product information here.
Comments on this entry are closed.
Hi james,
Iam very much impressed with your information. I kindly request you to help us.
Iam making use of SPSS Modeler 14 for predicting fraudulent customers by building a model using historical data(containing both customers who made fraud and not made fraud) and applied this model in an another stream along with new single customer. Finally, got predicted result along with probability value. I have used C5.0 Modeling algorithm(because it gave more accurate results). Is this way of predicting is correct? if not, how can i predict? Are there any JAVA API’s available for SPSS Modeler? How can i integrate this product with external products(like webservices / Java API’s…)? Please respond for this queries….
Thanks in Advance,
Ashok B.
In IBM – SPSS Modeler, you can build the model from a stream that includes all of your data prep functions, the model (C5 in your case), and assess the model to make sure it isn’t overfit and predicts accurately enough.
To deploy the model in Modeler, you can connect another data source to the stream, and as long as the data has the same fields as your original data used to build the model, you can run it through the exact same preprocessing nodes, through the C5 Gem, then export the record ID and model score to the database or flat file. If the new data has fields that are named differently from your original data, that’s OK–a filter node can change the names to the ones you need (in Modeler, exact name matching is used, so you have to the new data named the same as you used in building the models, case sensitive).
I’m not familiar with the latest IBM Modeler options now, but deployment depends on what system you want to deploy in. Is it a real-time transactional system, do you need to do batch updates daily, weekly, monthly, quarterly,… You can search the IBM – SPSS site for what kind of APIs are supportable: http://www.spss.com/software/modeler/
Dean