At PayPal, our team is tasked with seemlessly incorporating Responsible AI principles and practices into the company model development culture. One of the key component in Responsible AI is the Transparency and Explainability of AIs.
Think of why your credit card was declined when you were innocently purchasing a hotdog at the Knicks game, for example. Credit Card declines are usually automated with AI today. What led the AI to make such decision? Was it my FICO score? Did I not pay the bills on time? No, it can't be those.... I made sure to have a good credit score and paid off all my bills. Then what could it be?
As PayPal customer, I want to know these reasons, especially when automated decisions by PayPal is seemingly unreasonable. By providing why such actions were made by our automated models, or adverse action codes in other words, customers are left less baffled than before.
We achieve this by using SHAP (Shapley Values) to obtain feature importance of customer fraud detection models. SHAP is an industry standard methodology for extracting feature importance of predictive models. With SHAP, we are now able to assess blackbox models' (neural network) feature importance.
Assessment
We compare existing feature importance method to SHAP on the top 10 important feature by applications.