Deep water: Predicting water meter failures through a human-machine intelligence collaboration

Abstract

Take an AI learning algorithm and a human trainer with an experience in machine intelligence. Take piles of data, in the form of labeled examples. If you think that the task of training a machine that makes accurate decisions is easy as pie, you could not be further from reality. That is what has been missing from this story: getting off on the right foot. For a good start, crucial is the value of data which come large in quantity but low in quality. Beyond how we design our AI, fundamental is making our data valid for learning. We report here our experience in the creation of highly accurate training examples, based on the idea of filtering out all the impurities from a dataset containing 15 million water readings, provided by an Italian water supply company. This was accomplished allowing a human-machine collaboration, down to the implementation of an AI model capable to predict water meter failure.