Explainable Activity Recognition for Smart Home Systems

Smart home environments are designed to provide services that help improve the quality of life for occupant via a variety sensors and actuators installed throughout space. Many automated actions taken by smart governed output an underlying activity recognition system. However, systems may not be perfectly accurate therefore inconsistencies in operations can lead users reliant on predictions wonder "why did do that?" In this work, we build insights from Explainable Artificial Intelligence (XAI) techniques introduce explainable framework which leverage leading XAI methods generate natural language explanations explain what about led given classification. Within context remote caregiver monitoring, perform two-step evaluation: (a) utilize ML experts assess sensibility explanations, (b) recruit non-experts two user monitoring scenarios, synchronous asynchronous, effectiveness generated our framework. Our results show approach, SHAP, has 92% success rate generating sensible explanations. Moreover, 83% sampled scenarios preferred over simple label, underscoring need systems. Finally, some lose confidence accuracy model. We make recommendation regarding existing method leads best performance domain automation, discuss range topics future work further recognition.