Parameter Free Piecewise Dynamic Time Warping for time series classification

Vanel Steve Siyou Fotso (1) , Engelbert Mephu Nguifo (1) , Philippe Vaslin (1)

1 LIMOS - Laboratoire d'Informatique, de Modélisation et d'optimisation des Systèmes

Several improvements have been done in time series classification over the last decade. One of the best solutions is to use the Nearest Neighbour algorithm with Dynamic Time Warping(DTW), as the distance measure. Computing DTW is relatively expensive especially with very large time series. Piecewise Dynamic Time Warping (PDTW) is an efficient variant which consists of segmenting time series into fixed-length segments. However, the choice of the optimal size (or number) of segments remains a difficult challenge for end users. The Brute-force solution, a naive solution, repeats the classification with each segment size, and selects the one with the best accuracy. This solution is not appropriated especially when dealing with massive and large time series data. In this work, we propose a parameter free approach for PDTW, that finds the size (or number) of segments to be used with the Nearest Neighbour algorithm. Our approach is a heuristic that is parameter free since it does not require any domain specific tuning. Several properties of our heuris-tic are studied, and an extensive experimental comparison demonstrates its efficiency and effectiveness, in terms of accuracy and runtime.

Domaines: Apprentissage [cs.LG]; Intelligence artificielle [cs.AI]

Vanel Steve Siyou Fotso, Engelbert Mephu Nguifo, Philippe Vaslin. Parameter Free Piecewise Dynamic Time Warping for time series classification. 2016. ⟨hal-01408944⟩