DTW Barycenter Averaging
Dynamic Time Warping Barycenter Averaging · Also known as: DBA, DTW-BA, Barycenter Averaging
DTW Barycenter Averaging (DBA) is a method for computing the average or representative sequence of a set of time series that respects temporal warping and elastic distance. Unlike Euclidean averaging which requires point-wise alignment, DBA minimizes the sum of Dynamic Time Warping (DTW) distances, producing a meaningful average for sequences with flexible temporal alignments. Introduced by Petitjean and colleagues in 2011, it is widely used in time-series clustering and summarization.
Read the full method
Sign in with a free account to read this section.
Method map
The neighbourhood of related methods — select a node to explore.
When to use it
Use DBA when you need a representative sequence for a collection of time series with variable lengths or temporal distortions. It is essential for time-series clustering based on DTW distance, for summarizing multivariate sensor data, or for creating template sequences in shape-based retrieval tasks. DBA is superior to Euclidean averaging when temporal alignment matters.
Strengths & limitations
- Produces a meaningful average that respects temporal flexibility and elastic distance
- Works with variable-length sequences and complex temporal patterns
- Provides consistent clustering results when used with DTW-based distance metrics
- Robust to local time warping and speed variations in the data
- Computationally expensive—computing DTW distances is O(N^2) and the averaging process requires many iterations
- Non-convex optimization problem with multiple local minima depending on initialization
- The length of the output average sequence must be predetermined or estimated
- Can be sensitive to outliers or atypical sequences in the input set
Frequently asked
How is DBA different from simply taking the mean of time series at each time point?
Pointwise averaging (Euclidean mean) assumes all sequences are temporally aligned, which is unrealistic. DBA allows temporal warping—each input sequence can be stretched or compressed to match the average, producing a result that better captures the actual pattern.
What should the length of the output average sequence be?
The length is typically set to the average or median length of the input sequences, or can be estimated using heuristics. Some methods optimize the length jointly with the sequence values, though this increases computational cost.
How long does DBA computation take for a large dataset?
DBA involves K DTW distance computations per iteration, where K is the number of input sequences. Each DTW is O(N^2) for sequences of length N. For large datasets, use fastDTW approximations or limit the constraint window to reduce cost.
Can DBA handle sequences of very different lengths?
Yes, DBA naturally handles variable-length sequences through DTW. However, having very heterogeneous lengths (e.g., one sequence 10x longer than others) can produce poor averages—consider clustering by length or filtering outliers first.
Is DBA invertible or can I reconstruct original sequences from the average?
No, DBA is not invertible. The average sequence is a summary statistic. However, the DTW alignment paths computed during DBA can reveal how each input sequence aligns to the average, useful for downstream analysis.
Sources
- Salvador, S., & Chan, P. (2004). FastDTW: Toward accurate dynamic time warping in linear time and space. Intelligent Data Analysis, 11(5), 561–580. link ↗
- Petitjean, F., Ketterlin, A., & Gançarski, P. (2011). A global averaging method for dynamic time warping, with applications to clustering. Pattern Recognition, 44(3), 678–693. DOI: 10.1016/j.patcog.2010.09.013 ↗
- Cuturi, M., & Blondel, M. (2016). Soft-DTW: A differentiable loss function for time-series. arXiv preprint arXiv:1703.01541. link ↗
How to cite this page
ScholarGate. (2026, June 3). Dynamic Time Warping Barycenter Averaging. ScholarGate. https://scholargate.app/en/time-series/dtw-barycenter-averaging
Which method?
Set this method beside its closest kin and read them side by side — the library lays the books on the table; the choice is yours.
- Discrete Wavelet TransformTime Series↔ compare
- Dynamic Time WarpingDecision-making↔ compare
- Hierarchical ClusteringMachine learning↔ compare
- K-meansMachine learning↔ compare