Designing for ML UX
When developing a user’s experience for a feature that uses machine learning, a high rate of failure does not necessarily disqualify a feature from production. An algorithm’s failure rate instead affects the way a technological solution reveals itself to the user.
Take face recognition in photography as an example: there are demonstrably many ways to suggest a familiar face to the end user. Some experiences are more effective than others, but not because of the interface’s design. The best experiences are ones that compliment an algorithm’s precision and recall most effectively.
Figure A
In figure A, we see a simplified breakdown of the relationship between precision and recall as an algorithm tries to identify pictures of me. Every result solves the problem to some degree. Some might assume that the only acceptable option is in the top right quadrant. However, the best experience doesn’t come from high precision and recall alone, but from matching the best experience to the algorithm’s performance. Let’s take a look at the same chart, but with a basic call to action in an imaginary photography app.
Each example UX works best for the quadrant that it’s in. Conversely, if we were to propose the high precision, high recall solution (green) to the low precision, low recall algorithm (red), the result would be undoubtedly frustrating for the end user.
The last variable that we need to consider when designing a feature that uses machine learning is the impact of that specific problem on the user’s life. A solution that could have an incredibly high impact on a user’s life might have more room for false positives than false-positives for a low-impact problem.
Imagine a sophisticated app called Asbestos Detect. You simply take a picture and it can positively identify asbestos present in the picture. In this example, the algorithm’s precision and recall are vitally important and even disqualifying for certain quadrants. It absolutely needs to err on the side of high recall, regardless of precision. Correctly identifying one instance of asbestos is worth incorrectly identifying a number of others.
Machine learning, like UX, is a tool to solve a problem. How effective it is and the way that it surfaces itself depends first and foremost on what it attempts to solve. Only through deep empathy for the user and a solid understanding of the algorithm’s performance can the best solution be offered.
