Mathematical Programming Models for Mean Computation in Dynamic Time Warping Spaces

TL;DR

This paper develops polynomial-time bounds and mixed-integer nonlinear programming formulations to solve the NP-hard DTW-Mean problem in time series analysis, enabling more effective optimization approaches.

Contribution

The paper introduces new polynomial-time bounds and multiple MINLP formulations for the DTW-Mean problem, advancing beyond heuristic methods.

Findings

Formulations yield good results in specialized settings.

Polynomial-time bounds provide useful lower bounds.

Computational experiments compare multiple approaches.

Abstract

The dynamic time warping (dtw) distance is an established tool for mining time series data. The DTW-Mean problem consists of computing a series which minimizes the so-called Fr\'echet function, that is, the sum of squared dtw-distances to a given sample of time series. DTW-Mean is NP-hard and intractable in practice. So far, this challenging problem has been solved by various heuristic approaches without any performance guarantees. We give a polynomial-time algorithm yielding lower bounds on the domain of a mean time series which translate into lower bounds on the Fr\'echet function. We then formulate the problem as a discrete nonlinear optimization problem based on network flows. We introduce several mixed-integer nonlinear programming (MINLP) formulations in order to solve DTW-Mean optimally. Our formulations are based on techniques such as outer approximations and nonlinear reformulations of the well-known big M indicator constraints. Finally, we conduct several computational experiments to compare the different formulations on several instances derived from the UCR Time Series Classification Archive. While in general DTW-Mean remains quite challenging, our fomrulations yield good results in several important specialized probem settings.