Innovations in explainable AI in the context of real-world business applications

Description

In 1992, FICO introduced Falcon, a neural network-based fraud detection system that detects fraudulent payment card transactions in real time, which led to a dramatic reduction in payment card fraud across the globe. To assist analysts taking action on the scores generated by the model, FICO introduced the Reason Reporter, which provides reasons associated with the neural network scores—one of the first commercially available instances of explainable AI (XAI).

Scott Zoldi explores the algorithm and explains how it provides the explanations. Scott also offers a comparative study with local interpretable model-agnostic explanations (LIME)—a perturbation-based approach—to demonstrate the quality of the explanations as evaluated by human domain expert and outlines the theoretical underpinnings of why perturbation-based approaches fail to work well in many scenarios. Scott then digs deeper to show why looking for global score support in the historical data is a superior method for providing explanations. Understanding causality while providing model explanations have been elusive so far, as it requires ability to understand and parse latent features that actually drive the score. Scott concludes by briefly exploring LENNS (latent explanations neural network scoring), a patent-pending architecture that exposes more of what’s driving the score.