Prediction of Human Actions in a Smart Home Using Single and Ensemble of Classifiers

Abstract

There is a growing need for home automation systems that monitor and control a smart home environment to produce an efficient system that accurately predicts the needs of the human occupants. Past research has focused on the accuracy of prediction of a users future action. However, much of that work uses synthetic datasets which do not always reflect the real-world interactions that occur between an individual and the home environment. In addition, a focus on prediction accuracy often comes at the cost of slower processing time. This paper focuses on the prediction of future human actions in an intelligent environment with the goal of achieving both high prediction accuracy and response times that are appropriate for a real-time application environment. We performed experiments using the MavPad dataset, which was gathered from a fully-instrumented home environment and compared several different machine learning algorithms that included both single and ensemble classifiers. This study investigates whether an ensemble approach will satisfy the condition of realtime application much better than the performance of a single classifier. The results show that using a Support Vector Machine as a single classifier approach achieves the best results when using a group of sensors within a local zone, while the Random Forest classifier as an ensemble classifier approach achieves a higher performance when using sensors that are distributed across all zones inside the environment. The results lead us to the conclusion that dividing the environment into smaller zones assures the best performance of machine learning algorithms which is represented by the combination of maximum accuracy with a minimum time response for the prediction process.