Data By the Bay

Software UI and UX design have a few decades of established principles that guide designers in their tasks -- fundamental ideas like predictability, affordance, and graceful recovery from user errors. I have spent almost ten years trying to merge these ideas into the design of novel ML (and ML-assisted) products. Many of the traditional ideas become difficult to apply when you're intrinsically working with poorly understood data. What does it mean for an ML system to be predictable to the user, if the user has a poor understanding of the underlying data? If the ML system is not at least a little bit surprising, it's not doing its job. But how can a user feel confident if a surprising bit of feedback from the ML system is consistent with their own understanding of the world? How can they feel trust in the process when the process is itself usually beyond their understanding? Beyond that, while we have many years of experience on the "psychophysics" of visual perception, we have much less when it comes to user perception of probabilistic or statistical events. What little research we have is often contradictory, e.g. whether people better understand raw probabilities vs odds ratios vs qualitative statements like "very likely / likely / unlikely / etc.". I have spent many years studying people's in-the-field understanding and perceptions of probability in an applied setting, and this has come to shape much of how I design systems today. In many cases, I have found that these concerns motivated changes not just in a product's UX design, but changes in the underlying algorithms themselves. I'll be talking about highlights and best practices from this experience, spanning work from tiny startups and multiple Fortune 50 companies.