A General Approach for Supporting Time Series Matching Using Multiple-Warped Distances

Rodica Neamtu, Ramoza Ahsan, Cuong Nguyen, Charles Lovering, Elke A. Rundensteiner, Gábor N. Sárközy

Abstract: Time series are generated at an unprecedented rate in domains ranging from finance, medicine to education. Collections composed of heterogeneous, variable-length and misaligned times series are best explored using a plethora of dynamic time warping distances. However, the computational costs of using such elastic distances result in unacceptable response times. We thus design the first practical solution for the efficient <underline xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">GEN</u> eral <underline xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">EX</u> ploration of time series leveraging multiple warped distances. GENEX pre-processes time series data in metric point-wise distance spaces, while providing bounds for the accuracy of corresponding analytics derived in non-metric warped distance spaces. Our empirical evaluation on 66 benchmark datasets provides a comparative study of the accuracy and response times of diverse warped distances. We show that GENEX is a versatile yet highly efficient solution for processing expensive-to-compute warped distances over large datasets, with response times 3 to 5 orders of magnitude faster than state-of-art systems.