Intrinsic Properties of Human Accelerometer Data for Machine Learning

Time series data is often processed to extract features which better explain the sources and structure of the data. However, these processes make underlying assumptions about the nature of the time series. Two important intrinsic properties are the linearity and stationarity of the data. The large corpora on time series analyses include domains of economics, physics and engineering – thus cross domain approaches can yield useful insights into the data. Here we look at data from accelerometers, an important class of sensors. We employ widely used time series tests to provide novel analyses to establish their linear and stationary structure. This provides useful insights into the underlying processes which are being sensed and guide the type of temporal features, any preprocessing needed and suitable analyses to be performed. We briefly mention the use of this in a machine learning application.