WEASEL 2.0 is a dictionary-based time series classifier that combines dilated sliding windows with a randomised hyperparameter ensemble and a fixed-size dense feature representation. Two of its hyperparameter choices, the maximum ensemble size and the maximum window size, are specified by simple thresholding rules whose chosen thresholds are not empirically justified in the original paper. In this work we reproduce WEASEL 2.0 on 114 UCR datasets, achieving a mean accuracy of 0.865 and median of 0.928, closely matching the published values (Wilcoxon signed-rank, p = 0.655). We then test the sensitivity of four design choices: the downstream classifier, the absence of feature weighting, the maximum window-size rule, and the maximum ensemble-size rule. The first three are robust to perturbation. The fourth is over-provisioned for long-series datasets, motivating an adaptive rule that sets the maximum ensemble size from series length and number of classes. Evaluated on fixed-length datasets, the adaptive rule reduces peak fit memory by a median of 37 MB (mean 395 MB) and fit time by a median of 0.4 s (mean 4 s), with a median accuracy change of 0% (mean -0.11%). Memory and time savings concentrate on long-series datasets where the original rule allocates the largest ensemble size.
Figures & tables
Paper
Ours
Mean accuracy
0.863
0.865
Median accuracy
0.927
0.928
Datasets tested
114
114
Total fit time
1.71 h
0.57 h
Total predict time
1.79 h
0.69 h
Table 1: WEASEL 2.0 reproduction vs. the paper.
Classifier
Type
Mean
Median
N
WEASEL 2.0
D
0.865
0.928
114
MultiRocket
K
0.867
0.918
114
MiniRocket
K
0.861
0.909
114
ROCKET
K
0.857
0.908
114
R-DST
S
0.866
0.904
114
Hydra
K/D
0.858
0.900
113
Table 2: Reproduced accuracy summary for all classifiers (114 UCR datasets). D = Dictionary, K = Kernel, S = Shapelet.
Figure 1: Reproduced pairwise accuracy comparisons: WEASEL 2.0 (x-axis) vs. each competitor (y-axis). Points below the diagonal favour WEASEL 2.0.
Figure 2: Reproduced train, predict, and total runtime distributions (log scale) across fixed-length UCR datasets.
Figure 3: Reproduced trade-off plots between median accuracy and total runtime across 114 UCR datasets. The upper-left corner is ideal (fast and accurate). WEASEL 2.0 achieves the highest median accuracy with moderate runtime.
SGD
TF-IDF
Sublinear TF
Mean
Median
Mean
Median
Mean
Median
Fit time (s)
+8.19
+1.33
+40.74
+1.69
+42.90
+1.68
Predict time (s)
−0.79
−0.39
−0.20
−0.26
−0.16
−0.23
Peak memory (MB)
+0.55
+0.07
+43.20
+0.07
+42.72
+0.06
Accuracy (%)
−6.9
−3.4
−0.8
−0.3
+0.2
0.0
Table 3: Mean and median differences versus the WEASEL 2.0 baseline for the SGD, TF-IDF, and Sublinear TF variants on 22 UCR datasets. Bold indicates improvement over the baseline.
Halved
Adaptive
Mean
Median
Mean
Median
Fit time (s)
−3.64
−0.88
−3.99
−0.43
Predict time (s)
−4.15
−0.71
−3.41
−0.32
Peak memory (MB)
−306.4
−51.9
−394.7
−37.4
Accuracy (%)
−0.22
0.00
−0.11
0.00
Table 4: Mean and median differences versus the WEASEL 2.0 baseline for the halved-ensemble and adaptive-ensemble variants. Bold indicates the better of the two variants in each row. All differences are statistically significant ( p<0.05 ).
Figure 4: Trade-off plots between median accuracy and total runtime across 112 UCR datasets. The upper-left corner is ideal (fast and accurate). Comparison between WEASEL 2.0, adaptive, and halved variants.
Figure 5: Pairwise comparison of the adaptive and halved variants against WEASEL 2.0. A point below the diagonal indicates that WEASEL 2.0 is more accurate than the variant on that dataset.
Figure 6: Critical difference diagram on mean ranks across the UCR datasets. Methods connected by a horizontal bar are not significantly different under the Nemenyi test at α=0.05 .
Figure 7: Performance comparison for the new variants versus the WEASEL 2.0 baseline across five metrics on the UCR benchmark.