Empirical Identification of Non-stationary Dynamics in Time Series of Recordings

Non-stationarity time series are very common in physical, biological and in real-world systems in general, ranging from geophysics, econometrics or electroencephalography to logistics. Identifying, detecting and adapting learning algorithms to non-stationary environments is a fundamental task in many data mining scenarios; however it is often a major challenge for current methodologies. Data analysis in the context of time-varying statistical moments is a very active research direction in machine learning and in computational statistics; but theoretical insights into latent causes of non-stationarity in empirical data are very scarce. In this study, we evaluate the capacity of the trajectory classification error statistic in order to detect a significant variation in the underlying dynamics of data collected in multiple stages. We analysed qualitatively the conditions leading to observable changes in non-stationary data generated by Duffing nonlinear oscillators; which are ubiquitous models of complex classification problems. Analyses are further benchmarked in a dataset consisting of atmospheric pollutants time series.