Highly comparative time-series analysis
by Dr. Ben Fulcher
Abstract: An ongoing interdisciplinary research effort is focused on developing and applying time-series analysis methods to understand structure in real-world signals, from rainfall patterns and share prices through to human heart rates. The result is a huge array of potential techniques in the time-series analysis toolkit, but it remains unclear which of these myriad methods are best for a given application. In this talk, I will introduce a highly comparative approach to time-series analysis that compares across thousands of scientific time-series analysis methods (including information theoretic and entropy measures, stationarity estimates, scaling and correlation structural properties, and the performance of a range of time-series models). The approach automatically yields interpretable features that most accurately capture useful structure in a dataset, and has been successfully applied to classification and regression problems using speech, EEG, ECG, fMRI, seismic data, as well as a range of other data types. I will outline the types of scientific questions that can be answered using this approach and describe the available software package for performing this type of analysis (cf. www.comp-engine.org/timeseries).