The Role of AI Model Governance

A few weeks back I had the pleasure of participating in a LinkedIn Live broadcast with Scott Zoldi, FICO’s Chief Analytic Officer on the topic of model governance & artificial intelligence. I always welcome opportunities to participate in panel discussions with Scott (a genuine trailblazer in the field of AI and predictive analytics), as I invariably learn something new about the field of AI every time I do!

In this particular session, we covered the topic of the role of model governance and AI, particularly through the lens of the FICO® Score that my team is responsible for developing.  Here are highlights of some key points made during the discussion:

• Many treat Artificial Intelligence and Machine Learning as having some sort of mystical ability…
  • At the end of the day, it is just another (powerful) analytic technique which requires proper oversight, problem definition, safe guards, expertise, and model governance to ensure that the resulting model is robust and reliable.
• AI has been used extensively within FICO for more than 30 years
  • Our fraud modeling group pioneered much of our work in the AI space, obtaining some of the earliest patents of the use of neural nets to detect fraudulent credit card transactions.  And that is because the fraud use case is very well suited to AI, which drives highly predictive models with low false positive rates even in a setting of significant ‘class imbalance’ between the fraud/non-fraud outcomes being predicted. 
  • The FICO® Score is under intense scrutiny---It is used by 90% of the top U.S. lenders for credit decisions each year.  In that context, where transparency, explainability and palatability are so critical, we leverage traditional scorecard technology to build our models (although we are continually monitoring progress within interpretable machine learning models), and utilize AI as a complementary tool for deriving fast insights into the data rather than as the underlying methodology for arriving at the final model weights.
• Garbage in = Garbage out
  • No modeling technology including AI can overcome issues with data quality and bias, which is why the lion’s share (over 2/3rds) of the timeline and resource hours devoted to a FICO® Score redevelopment is actually associated with the up front processing, QA, and cleansing of the redevelopment data sample.
• Document, document, document!
  • When it comes to good model governance, ensuring complete and robust model documentation is very important.  And not just documenting the roads taken and the rationale for those decisions, but also some of the other options that were evaluated that ended up not being pursued, and the reasoning for that. The FICO^® Score’s longevity really has underscored the importance of documentation of analyses and decisions made during the score development process.  You don’t want to find yourself having to recreate or justify an analysis/decision made years ago to a regulator because you didn’t sufficiently document or capture the right analysis at time of development.

You can watch the full conversation between Scott and me on topic of model governance & artificial intelligence. And don’t forget to follow Dr. Scott Zoldi on LinkedIn and Twitter @ScottZoldi to keep up with his latest thoughts on AI, data science and more.