Monitoring Routine Changes From Daily Smartphone Usage

Abstract

The detection of user routine changes from smartphone sensor data is investigated in this study. A smartphone application is used to record multi-modal sensor data. A dataset from 60 users was used for activity classification. Anomaly detection was performed on these activities to detect and characterise abnormal behavioural changes. A Multi-task Multilayer Perceptron Neural Network was used for activity classification. Four different anomaly detection architectures were compared, using two weeks of data for training. An accuracy of 65.7 percent was achieved for activity classification of the 14 most common human activities. A One-class Support Vector Machine yielded the best results for the anomaly detection, with an accuracy of 76.8 percent. These preliminary results show a potential of the proposed methods to detect and characterise changes in human routine.