[Paper Review] SparseDTW: A Novel Approach to Speed up Dynamic Time Warping

Ghazi Al‐Naymat, Sanjay Chawla+1 more|arXiv (Cornell University)|Jan 13, 2012

Time Series Analysis and Forecasting19 references131 citations

TL;DR

SparseDTW proposes a space-efficient, optimal Dynamic Time Warping (DTW) algorithm that dynamically adapts to inherent similarity and correlation between time series, reducing memory usage without sacrificing optimality. Unlike banding or indexing methods, it uses sparse matrix representation to compute only relevant cells, achieving significant speedups and memory savings while guaranteeing optimal alignment.

ABSTRACT

We present a new space-efficient approach, (SparseDTW), to compute the Dynamic Time Warping (DTW) distance between two time series that always yields the optimal result. This is in contrast to other known approaches which typically sacrifice optimality to attain space efficiency. The main idea behind our approach is to dynamically exploit the existence of similarity and/or correlation between the time series. The more the similarity between the time series the less space required to compute the DTW between them. To the best of our knowledge, all other techniques to speedup DTW, impose apriori constraints and do not exploit similarity characteristics that may be present in the data. We conduct experiments and demonstrate that SparseDTW outperforms previous approaches.

Motivation & Objective

• Address the high space complexity of standard DTW, which scales as O(mn) and limits its use on long time series.
• Overcome the trade-off in existing speed-up methods between efficiency and optimality, where constraints or abstractions sacrifice accuracy.
• Develop a method that adapts to data characteristics—specifically similarity and correlation—without requiring apriori assumptions.
• Enable practical computation of DTW on large-scale time series data by minimizing stored matrix cells while preserving optimality.
• Provide a framework compatible with lower-bound indexing techniques, enhancing performance in similarity search workloads.

Proposed method

• Dynamically construct a sparse representation of the DTW warping matrix based on observed similarity and correlation between time series.
• Use a sparse matrix data structure to store only the cells that are potentially part of the optimal warping path, avoiding full O(mn) storage.
• Apply dynamic programming principles to compute DTW distances only over the sparse set of relevant cells, reducing both time and space complexity.
• Evolve the search band adaptively during computation, unlike fixed-bands in Sakoe-Chiba or Itakura constraints, ensuring no loss of optimality.
• Leverage the fact that highly correlated sequences have warping paths close to the diagonal, thus minimizing the number of cells to evaluate.
• Integrate with lower-bound filtering techniques (e.g., LBF) since the method guarantees optimal results, enabling efficient similarity search pipelines.

Experimental results

Research questions

• RQ1 Can we reduce the space and time complexity of DTW without sacrificing optimality by exploiting inherent data similarity?
• RQ2 How does the adaptive sparsity of the warping matrix compare to fixed-banding approaches in terms of memory usage and accuracy?
• RQ3 To what extent does the correlation between time series influence the number of cells opened during DTW computation?
• RQ4 Can a sparse DTW approach be efficiently combined with lower-bound filtering techniques used in time series similarity search?
• RQ5 Does the dynamic adaptation of the warping band lead to consistent performance improvements across diverse real-world and synthetic datasets?

Key findings

• SparseDTW consistently outperforms standard DTW, BandDTW, and the Divide-and-Conquer (DC) method in both runtime and memory usage across all tested datasets.
• For the GunX dataset, SparseDTW reduced the number of computed cells from 75,076 (DTW) to 17,220, achieving a 77% reduction in cell computation.
• In the Burst-Water dataset, SparseDTW computed only 951,150 cells compared to 2,190,000 for standard DTW, a 56% reduction in cell computation.
• SparseDTW achieved optimal results in all cases, while BandDTW showed errors ranging from 30% to 500% compared to standard DTW.
• The algorithm’s performance improves significantly with higher correlation: sequences with strong similarity required far fewer open cells than uncorrelated ones.
• For datasets larger than 6,000 points, standard DTW became infeasible due to memory constraints, while SparseDTW remained practical and efficient.