Transparent Machine Learning in IBM Business Automation
How to use ML in critical and regulated business decisions.
It’s becoming increasingly common to use Machine Learning (ML)to make more precise and personalized B2C decisions. However, as we’ve explored in in this introduction , the ability to explain and justify such decisions is becoming increasingly important. It’s driven both by customer expectations, business controls and government regulations.
In Automation Decision Services (ADS) — a component of Cloud Pak for Business Automation — there’s been support for invoking remote ML models for a couple of years. This low-code capability supports flexible invocation of models, including models where an explanation is returned alongside the prediction.
That said, while we can to some degree explain individual predictions from black-box ML models, how about building and using ML models that are inherently transparent? In other words, an ML model that is readable by a business user?
In the latest releases of Automation Decision Services (ADS) it is now possible to import two types of transparent ML models — Rule Sets and ScoreCards. Using rule sets as a transparent model type to make predictions is very natural in ADS, since ADS is build around authoring and managing If-Then rules. Similarly, ScoreCards can be naturally represented in ADS as a table for each subscore.
The overall flow for learning and using rule sets/scorecards inside decisions is essentially the same as a regular ML training pipeline, with the main exception being the last step where we import the ML models as a set of rules/tables into ADS, instead of deploying the ML model behind a REST API and making a remote invocation:
To facilitate this flow, a set of tools and industry standards are being chained together:
- Rule Learning algorithms (developed by IBM Research and open-sourced in the AI Explainability 360 toolkit with sample Jupyter Notebooks) and also the RuleLearn repository.
- As an alternative to using notebooks, there’s an interactive Rule Learning UI called AIMEE on top of the algorithms (not yet productized, but available to IBM sales/consultants for client engagements)
- Ability to export rule sets in PMML RuleSet and PMML ScoreCard format (an industry standard managed by the Data Mining Group consortium)
- Capability to import PMML RuleSets/ScoreCards into either IBM Automation Decision Services (native feature since 22.0.1 release) or alternatively importing into IBM Decision Manager Open Edition (formerly called RedHat Decision Manager, which now supports rule sets through new feature in Drools open source project and ScoreCards with a native editor in KIE Sandbox)
When should you use remote black-box ML models, and when is transparent rule sets preferable? Depends on the use case and the requirements. The comparison table below summarizes the pros and cons with each approach. As a summary, for critical decisions where business ownership and regulations are important, transparent ML models are clearly better. Conversely, for less-scrutinized predictions based on complex ML models owned by the data science team, go for remote ML models.
Additional pointers:
Greger works for IBM and is based in France. He’s a product manager and researcher, focusing on Machine Learning for Business Automation.