DTW search stops using one ruler for every pair: WETW predicts per-pair weights and runs 40 times faster than the UCR Suite on long series
Synopsis
WETW changes the comparison rule in DTW search from a fixed distance to weights generated per series pair: two fixed-length embeddings go into a small symmetric network that predicts one weight per embedding dimension, without computing an alignment path. On eight datasets, the WETW-FV pipeline verifies a shortlist with exact DTW and reaches a mean recall@10 of 0.93 at 11.4 times the speed of the fastest exact method; on 48K-point series it stays 40 times faster than the UCR Suite at a recall@10 of 0.91.
Interpretation
Time-series similarity search can now let the comparison rule vary with the series pair instead of applying one fixed distance to every pair of embeddings. Earlier learned DTW approximations mapped each series to an embedding and then compared every pair with the same fixed measure, usually Euclidean distance, whereas DTW itself chooses a different alignment for each pair. With the same frozen WESee embeddings and the same supervision budget, WETW reaches a mean relative error of 0.04 against 0.08 for the full-matrix static measure and 0.24 for Euclidean distance, and a mean recall@1 of 0.79 against at most 0.73 for the static measures.
The search pipeline first ranks all database embeddings by cheap Euclidean distance, then rescores the closest part with WETW, and finally runs exact DTW only on a small shortlist. Exact methods rely on lower-bound pruning, yet every candidate that survives a bound still needs a full DTW computation; on oscillating signals most query points fall inside the candidate envelope, so the bounds prune little. Across the eight datasets, WETW-FV reaches a recall@10 of 0.95 within a 10% shortlist and is the fastest filter at that recall on every dataset; on average it reaches a mean recall@10 of 0.93 at the speed of the fastest exact search, while the best learned baseline reaches 0.59 at the same query time.
Perspective
The evaluation covers whole-series k-nearest-neighbour search over fixed-length univariate series of 256 to 48K points, with databases of 20.9K to 840K series, using Sakoe–Chiba DTW with a fixed fractional window; query times are measured on a single CPU thread, and NeuralWarp scoring runs on an A100 GPU because it is too slow on CPU.
Regression alone already gives a low mean relative error yet a recall@1 of only 0.79, so fitting distances does not guarantee the order of the closest neighbours; on AudioMNIST-DS and FruitFlies a static measure retrieves the nearest neighbour slightly more often than WETW. Multivariate and variable-length series, cross-dataset transfer and behaviour under distribution shift are not evaluated.
