Machine learning algorithms for the activity monitoring of elders by home sensor network

Abstract

The aim of this research is to measure and cluster the activity of ageing people, through a minimally invasive home sensor network, realised by movement and light sensors, installed in seven apartments. Unsupervised machine learning (UML) algorithms are used to cluster daily data in different pattern of activation of light and movement sensors. Data are grouped in 6 time slots per day of 4 hours each and the activations of each sensor are counted. The chosen number of clusters is set to 2 and best clustering results are obtained for the first time slot of House 4 and House 5 with a Silhouette score of 0.7 and 0.9, suggesting clear separation between the days belonging to 2 different clusters. In addition, to provide an explanation to the clustering algorithm, a supervised machine learning (SML) algorithm is used to establish which sensor, in the different time slots of the day, is ruling the clustering. Moreover, a Decision Tree algorithm (DT) is used to understand the clustering methodology adopted by the UML algorithm for each time window of the different houses. DT allows identifying which sensor may cause the assignment of a diversity condition between the days of the dataset.