Reducing the intrusion of user-trained activity recognition systems

Many supervised activity recognition systems require a fully labelled time-series for accurate classification. However, gathering these labels is a difficult and often unrealistic task, especially over long-time frames or outside of laboratory conditions. A potential solution is through diary studies, allowing for a user-trained activity recognition system. Requests will be presented on the user's smart device and while this approach will be significantly less intrusive than current methods, frequent or inappropriately timed requests could reduce user acceptance. This paper proposes to further reduce user intrusion by making a prediction about the next user request and analyzing the classifiers confidence in this prediction. Two methods are presented, and with careful selection of the confidence threshold, they resulted in up to a 35% reduction in user requests with a minimal reduction in accuracy.