DrinkWatch

Abstract

We describe in detail the development of DrinkWatch, a wellbeing application, which supports (alcoholic and non-alcoholic) drink activity logging. DrinkWatch runs on a smartwatch device and makes use of machine learning to recognize drink activities based on the smartwatch»s inbuilt sensors. DrinkWatch differs from other mobile machine learning applications by triggering feedback requests from its user in order to cooperatively learn the user»s personalized and contextual drink activities. The cooperative approach aims to reduce limitations in learning performance and to increase the user experience of machine learning based applications. We discuss why the need for cooperative machine learning approaches is increasing and describe lessons that we have learned throughout the development process of DrinkWatch and insights based on initial experiments with users. For example, we demonstrate that six to eight hours of annotated real world data are sufficient to train a reliable base model.