Decision Management (ODM, ADS)

Posts in the series:

1. Benefits of Decision Modeling
2. What is a decision model and what is DMN
3. Decision Modeling: Decision Requirements
4. Decision Modeling: Decision Logic
5. Decision Modeling: Finding and modeling decisions
6. Input Data and Knowledge Sources
7. Building decision tables from decision models
8. ML, AI and other forms of Data Analysis and Advanced Analytics [this post]

--

There is tremendous interest in machine learning (ML) and artificial intelligence (AI) today – especially large language models (LLMs) or Generative AI (GenAI). One of the great things about decision modeling is that it has always supported the inclusion of these kinds of analytics. After all, the reason we invest in data and analytics or ML is to make better, more data-driven decisions. And what an analytic model does, is answer a question. It might do this directly in the case of a predictive or segmentation model - how likely is this customer to default on this loan in the next 12 weeks, say. Or it might do it indirectly in the case of “softer” analytics - a report might show the general distribution of the cost of a treatment that can be used to set boundaries outside which transactions must be reviewed. In fact, this is even true with advanced forms of mathematical optimization, which might decide on an optimal packing arrangement or schedule.

When an analytic or advanced model is executable on a single transaction, it makes a decision about that transaction and answers a question (in a defined way) about that transaction. This means that a decision in you model can be based on it – not on explicit prescriptive logic but on the result of this data analysis, machine learning or AI. These decisions won’t have decision tables or business user-driven logic. Instead, there will be an algorithm or formula. However, the result of this can be required by other decisions in the model just as though it was produced by rules, firmly integrating your advanced analytics into your decision-making. This works even when several distinct predictions must be combined to make a business decision – the decision model shows how the models are combined.

For instance, consider making a preventative offer to retain a customer. This decision, shown below, involves both rules-based decisions (shown in green) and machine learning-based decisions, predictions (show in orange). This decision model shows that deciding what preventative offer to make (if any), depends on the customer having an eligible renewal approaching, the likelihood that they will churn, whether they are worth keeping and the best offer to retain them. Drilling down, you can see that the decision as to which offer is “best” depends both on predictions of propensity (machine learning) and eligibility (rules-based).

If the analysis is softer, not executable, then some rules will need to be written based on the analysis. In this situation, the analytic work will show up as a knowledge source that acts as an authority for those rules, so we remember what analysis was done to create them in the first place.

In fact, we can use knowledge sources quite extensively to document our analytic work. Even when the end result is an executable model – a prediction or score – we can use knowledge sources to document how the work was done, what missing value imputation approach was taken, what feature distribution was expected etc. This brings documentation out of the python code, and into a visual blueprint where it can be tied to the decision being made and shown to those who must act on it or trust it.

MLOps – the operationalization of machine learning models – can also be supported. The features and calculated values that are fed into an ML model can be documented as decisions too, allowing them to be reused between ML models – and between ML and expert rules-based models. Business users are more likely to trust these models when more of the “bones” of their development are visible and explainable. And the use of a decision model and decision platform makes it possible to show exactly the ML is being used in every decision. The whole decision, including its ML and AI components, to be documented in a single, coherent model.

For instance, this example uses knowledge sources to document feature distribution, missing data imputation approaches and the modeling itself. It also uses decisions to show what features are being created to drive the model and another decision to decide it is time to retrain the model!

And you can transition this kind of model into your decisioning platform. Here, for instance, is a model in ADS that uses a machine learning model as part of its execution.

If you want more detail, you can get a text book on the approach (written by me and Jan Purchase): Real-World Decision Modeling with DMN 2nd Edition. If you or your company need help with decision modeling, drop me a line james@decisionmanagementsolutions.com and we can schedule a quick call to discuss. And if you’re excited about decision modeling and keen to do more, why not join DecisionAutomation.Org and participate?

#IBMOperationalDecisionManager#OperationalDecisionManager(ODM)#AutomationDecisionServices(ADS)#DecisionManagerOpenEdition#DMN#IBMChampions#decisionmgt#DecisionAutomation#IBMChampion