B1556
Title: Measuring non-stationarity in large time series: A spectral unsupervised learning approach
Authors: Sourav Das - Curtin University (Australia) [presenting]
Guillermo Cuauhtemoctzin Granados Garcia - King Abdullah Universe of Science and Technology (Saudi Arabia)
Hernando Ombao - King Abdullah University of Science and Technology (KAUST) (Saudi Arabia)
Abstract: EEG recorded during an epileptic seizure is an example of a typical, burgeoning, non-stationary time series data. Such data exhibit time-dependent changes in variance in the amplitudes of the various oscillating waveforms. The spectral density function $f(w)$ is the unique time-invariant signature of a second-order stationary time series. Motivated by the challenges of the seizure EEG data, a measure of second-order non-stationarity $R(t)$ is proposed using the conventional periodogram $I(w_k)$, an estimator for $f(w)$. $R(t)$ measures the deviation of the periodogram from second-order stationarity. Its utility is highlighted in monitoring disease incidence and aetiology.