Scaling up Human-AI Collaboration for the New Frontier of Payments Security
Presented at Data Science Salon Austin: Machine Learning in the Enterprise
This presentation is an introduction to using UX, UI, and cognitive science principles to help end users understand AI recommendations within the context of a case study of Q2's cybersecurity payments fraud product, Sentinel.
The Next Frontier of Identifying Payment Fraud: Checks
The Sentinel team works on many types of payments fraud, but the next big thing is going to be check fraud. The impact of the pandemic on banking practices has resulted in a huge increase in MRDC, or Mobile Remote Deposit Checks, as well as a corresponding increase in fraud since digital images of checks are easier and cheaper to fake than physical checks. A surprisingly large number, 40% of small business/commercial payments, is still conducted through checks, so this is a problem that is impacting a lot of people.
Information to Insight: Getting Data into the Machine, Transforming Data into Insights, Communicating Insights to Human Users
I worked with the Chief Data Scientist at Q2 to document the data science process and theory of how Sentinel works. We use a combination of cutting edge and simple, foundational data science methods, such as decision trees, clustering algorithm, and, in the future, computer vision for check signatures to identify patterns and anomalies in the data.
Human-AI Collaboration
All the data and insights in the world won't be useful if it can't be communicated and put in action by the people who need it, however. Using my CX and user research experience, I interviewed fraud specialists across our financial institution network to understand the commonalities and differences in their fraud detection workflows. These insights allowed me to redesign the suspect transaction report into a format that could more effectively communicate to these busy fraud experts as well as updating it to reflect the advances in decision-making algorithms that the data science team had created.
Fundamentally, data science and machine learning alone won't solve the payment fraud problem; humans need to take action on it. The software systems we design need to fit into the human systems that exist today.
Design Principles for Explainable AI (XAI)
Since this conference was oriented toward data scientists and ML product managers, I provided an introduction to UX, UI, and cognitive science principles that can be used in interfaces to explain AI decision-making. Each principle is paired with an example of how it is applied in the Sentinel suspect transaction report.
Creating a Fraud Feedback Loop to Scale the Power of Machine Learning Across the System
Finally, everything thus far has been focused on designing the system to communicate from the machine to the human user. The next step is creating a way for the human to give feedback to the machine learning system. This new idea of a true fraud database and decisioning workflow will allow Sentinel to take feedback from fraud specialists to improve its ML algorithms and recommendations. This will allow us to scale the power of the system across all financial institutions in the network, unlocking a new power of scale for machine learning in the enterprise.
