WETW learns pair-specific distances to speed up DTW time-series search
An arXiv paper proposes a learned distance that adapts to each pair of time series, paired with a filter-and-verify pipeline that stays fast while returning exact DTW distances.
In plain words
Searching large collections of time series for similar patterns is slow because the standard method, Dynamic Time Warping (DTW), compares every candidate in a costly way. WETW is a learned shortcut: a small network scores how similar two shortened series representations are, adjusting the comparison for each pair. A pipeline uses these scores to shortlist candidates, then checks the shortlist with exact DTW.
What was announced
The authors observe that once two series are aligned, DTW amounts to a Euclidean distance over coordinates chosen for that pair. Building on this, WETW uses a small symmetric network to predict one weight per embedding dimension and pair, without computing an alignment. The paper reports this halves the error of the best static distance on the same embeddings. The embeddings come from the WESEE encoder, whose multi-scale temporal feature extractor preserves timing information that global pooling removes, raising recall@1 from 0.90 to 0.95, though WETW works with any encoder.
The paper also argues that fitting DTW distances by regression alone is insufficient for search, since it misorders the closest neighbours; adding ranking regularization raises recall@1 from 0.79 to 0.95. The filter-and-verify pipeline, WETW-FV, uses WETW to build a shortlist and exact DTW to verify it, so returned distances are exact even though search is approximate. Across eight datasets, the authors report WETW-FV is 11.4x faster than the fastest exact method compared, at a mean recall@10 of 0.93, while the best learned baseline reaches 0.59 at the same speed. On 48K-point series, it remains 40x faster than the UCR Suite at recall@10 of 0.91.
What it does not show
- Evidence is limited to the paper's abstract; the full article with methods, datasets and experimental detail was not reviewed, so the reported speedups and recall figures could not be independently checked.
Sources
Source-support review completed 17 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 17 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.