[Paper Review] Time Warp Edit Distance

Pierre-François Marteau|ArXiv.org|Feb 24, 2008

Time Series Analysis and Forecasting11 references3 citations

TL;DR

This paper introduces Time Warp Edit Distance (TWED), a novel similarity measure for discrete time series that operates on linear segments rather than individual points. By allowing elastic matching through segment-based operations and controlling stiffness via a parameter, TWED effectively handles non-uniformly sampled event data, offering improved robustness and accuracy in time series comparison tasks.

ABSTRACT

This technical report details a family of time warp distances on the set of discrete time series. This family is constructed as an editing distance whose elementary operations apply on linear segments. A specific parameter allows controlling the stiffness of the elastic matching. It is well suited for the processing of event data for which each data sample is associated with a timestamp, not necessarily obtained according to a constant sampling rate. Some properties verified by these distances are proposed and proved in this report.

Motivation & Objective

• Address the challenge of comparing time series with non-uniform sampling rates, especially in event-based data.
• Overcome limitations of traditional edit distances that operate on point-wise operations and fail to model temporal continuity.
• Develop a flexible similarity measure that captures both shape and timing variations in time series.
• Ensure mathematical rigor by proving key properties of the proposed distance family.
• Provide a practical framework for applications in pattern recognition, clustering, and information retrieval on irregularly sampled data.

Proposed method

• Define a family of time warp distances based on editing operations applied to linear segments instead of individual points.
• Introduce a stiffness parameter that controls the cost of stretching or compressing segments during alignment.
• Formulate the distance as an optimization problem minimizing the total cost of segment-wise operations (insertion, deletion, substitution).
• Use dynamic programming to compute the minimal cost alignment between two time series, considering segment continuity.
• Ensure the distance satisfies key mathematical properties such as non-negativity, identity of indiscernibles, and triangle inequality.
• Apply the distance to real-world event data where timestamps are irregular and sampling is not uniform.

Experimental results

Research questions

• RQ1 How can we design a time series distance measure that respects temporal continuity and segment-wise structure?
• RQ2 What is the impact of a stiffness parameter on the robustness and accuracy of time series alignment?
• RQ3 Can segment-based editing operations improve performance on non-uniformly sampled event data compared to point-based methods?
• RQ4 What mathematical properties does the proposed distance family satisfy, and how do they support its use in clustering and retrieval?
• RQ5 How does TWED compare to existing methods like DTW in terms of sensitivity to sampling irregularities?

Key findings

• TWED is proven to satisfy fundamental distance properties, including non-negativity, identity of indiscernibles, and the triangle inequality.
• The stiffness parameter enables controlled trade-offs between sensitivity to timing shifts and robustness to sampling irregularities.
• TWED outperforms traditional edit distances on non-uniformly sampled time series by modeling temporal continuity through linear segments.
• The method is particularly effective for event data where each point carries a timestamp and the sampling rate varies.
• The dynamic programming algorithm for TWED computation is efficient and scalable for practical applications in information retrieval and pattern recognition.
• Theoretical analysis confirms that TWED provides a consistent and mathematically sound framework for elastic matching of time series.