HOW DO YOU KNOW IF 85% ACCURACY is good enough for your application?

Abstract

Many people who want to build mobile applications based on sensing systems have to decide what type of sensing approaches to use, often trading off factors like accuracy, time, and battery life. If you're building an app that uses location, motion data, or inferred data about usage, there will necessarily be inaccuracy in the output of those systems -- some error that cannot be completely eliminated from the system. In the ubiquitous computing community, there is an unofficial standard that 85% accuracy is "good enough" for sensing based on machine learning. But it's not so simple to say that 85% should be your target accuracy to consider a system useful. Maybe 85% accuracy is good enough for an app that helps you identify wasteful appliances so you can conserve water in your home, but is completely unacceptable when three times out of 20 your fall detection app accidentally calls emergency services thinking your loved one has fallen down even though she hasn't.