aeon-toolkit · TonyBagnall · Aug 31, 2026 · Aug 31, 2026 · Aug 31, 2026 · Aug 31, 2026
diff --git a/README.md b/README.md
@@ -23,6 +23,12 @@ The archive is described in
 Classification](https://arxiv.org/abs/2603.20352) (arXiv:2603.20352). If you use the
 archive, please cite it; `CITATION.cff` has the full entry.
 
+The wider literature the archive is measured against is reviewed in
+[Benchmark-Driven Multivariate Time Series Classification: The Role of the UEA MTSC
+Archive](https://eprints.soton.ac.uk/512258/), a survey of 142 papers using the UEA
+archive. The accuracies it collected are in
+[`published_results/`](published_results/), and are what our runs are compared against.
+
 The current paper version describes:
 
 - 133 unique MTSC problems
@@ -32,33 +38,34 @@ The current paper version describes:
 ### Multiverse-core leaderboard
 
 <!-- LEADERBOARD:START -->
-| # | Estimator | Accuracy | Balanced accuracy | AUROC | F1 | Log loss &darr; | Sensitivity | Specificity | Accuracy rank |
+| # | Estimator | Accuracy rank | Accuracy | Balanced accuracy | AUROC | F1 | Log loss &darr; | Sensitivity | Specificity |
 |---|---|---|---|---|---|---|---|---|---|
-| 1 | HC2 | **0.7909** | 0.7518 | **0.8990** | 0.7273 | **0.5383** | 0.7459 | **0.7943** | **7.64** |
-| 2 | MRHydra | 0.7837 | **0.7564** | 0.8105 | **0.7316** | 7.7974 | **0.7642** | 0.7757 | 8.14 |
-| 3 | RDST | 0.7734 | 0.7333 | 0.7912 | 0.6991 | 8.1667 | 0.7109 | 0.7874 | 9.02 |
-| 4 | RIST | 0.7720 | 0.7397 | 0.8748 | 0.7147 | 0.6218 | 0.7408 | 0.7655 | 9.60 |
-| 5 | DrCIF | 0.7747 | 0.7429 | 0.8813 | 0.7173 | 0.6484 | 0.7397 | 0.7708 | 9.86 |
-| 6 | FreshPRINCE | 0.7743 | 0.7487 | 0.8745 | 0.7211 | 0.6007 | 0.7414 | 0.7770 | 9.87 |
-| 7 | CIF | 0.7781 | 0.7471 | 0.8908 | 0.7212 | 0.6430 | 0.7441 | 0.7753 | 9.94 |
-| 8 | QUANT | 0.7720 | 0.7462 | 0.8831 | 0.7189 | 0.6175 | 0.7521 | 0.7581 | 10.29 |
-| 9 | Arsenal | 0.7680 | 0.7321 | 0.8457 | 0.7024 | 3.8631 | 0.7257 | 0.7732 | 10.41 |
-| 10 | ROCKET | 0.7690 | 0.7326 | 0.7925 | 0.7019 | 8.3249 | 0.7200 | 0.7764 | 10.58 |
-| 11 | LITETime-MV | 0.7506 | 0.7299 | 0.8513 | 0.6820 | 1.3206 | 0.7132 | 0.7637 | 10.92 |
-| 12 | STSF | 0.7724 | 0.7477 | 0.8804 | 0.7080 | 0.6432 | 0.7345 | 0.7826 | 11.29 |
-| 13 | H-InceptionTime | 0.7408 | 0.7190 | 0.8496 | 0.6838 | 1.3227 | 0.7223 | 0.7378 | 11.39 |
-| 14 | LiteTIME | 0.7341 | 0.7104 | 0.8394 | 0.6680 | 1.4776 | 0.7113 | 0.7336 | 12.08 |
-| 15 | PatchMTSC | 0.7428 | 0.6897 | 0.8261 | 0.6533 | 0.7655 | 0.6852 | 0.7352 | 12.77 |
-| 16 | ConvTran | 0.7462 | 0.7102 | 0.8592 | 0.6767 | 0.8190 | 0.7159 | 0.7345 | 12.89 |
-| 17 | Catch22 | 0.7475 | 0.7181 | 0.8697 | 0.6922 | 0.7147 | 0.7240 | 0.7374 | 12.93 |
-| 18 | STC | 0.7545 | 0.7172 | 0.8744 | 0.6940 | 0.6391 | 0.7185 | 0.7537 | 13.63 |
-| 19 | TSF | 0.7515 | 0.7236 | 0.8740 | 0.6883 | 0.7252 | 0.7093 | 0.7606 | 13.63 |
-| 20 | TDE | 0.7262 | 0.6813 | 0.8374 | 0.6382 | 0.8869 | 0.6714 | 0.7344 | 14.21 |
-| 21 | Summary | 0.6858 | 0.6574 | 0.8268 | 0.6230 | 0.9123 | 0.6574 | 0.6844 | 16.12 |
-| 22 | 1NN-DTW | 0.6712 | 0.6454 | 0.7197 | 0.6136 | 11.8506 | 0.6521 | 0.6636 | 17.82 |
-| 23 | Dummy | 0.3645 | 0.3029 | 0.5000 | 0.1507 | 1.4067 | 0.2855 | 0.3816 | 20.95 |
-
-Test results for default train/test split. All classifiers trained with default settings. Average over the 52 Multiverse-core datasets with results for every estimator on every metric, ordered by average accuracy rank. Best in each column in bold.
+| 1 | HC2 | **7.86** | **0.7909** | 0.7518 | **0.8990** | 0.7273 | **0.5383** | 0.7459 | **0.7943** |
+| 2 | MRHydra | 8.31 | 0.7837 | **0.7564** | 0.8105 | **0.7316** | 7.7974 | **0.7642** | 0.7757 |
+| 3 | RDST | 9.16 | 0.7734 | 0.7333 | 0.7912 | 0.6991 | 8.1667 | 0.7109 | 0.7874 |
+| 4 | RIST | 9.76 | 0.7720 | 0.7397 | 0.8748 | 0.7147 | 0.6218 | 0.7408 | 0.7655 |
+| 5 | DrCIF | 10.07 | 0.7747 | 0.7429 | 0.8813 | 0.7173 | 0.6484 | 0.7397 | 0.7708 |
+| 6 | FreshPRINCE | 10.10 | 0.7743 | 0.7487 | 0.8745 | 0.7211 | 0.6007 | 0.7414 | 0.7770 |
+| 7 | CIF | 10.13 | 0.7781 | 0.7471 | 0.8908 | 0.7212 | 0.6430 | 0.7441 | 0.7753 |
+| 8 | QUANT | 10.55 | 0.7720 | 0.7462 | 0.8831 | 0.7189 | 0.6175 | 0.7521 | 0.7581 |
+| 9 | Arsenal | 10.70 | 0.7680 | 0.7321 | 0.8457 | 0.7024 | 3.8631 | 0.7257 | 0.7732 |
+| 10 | ROCKET | 10.82 | 0.7690 | 0.7326 | 0.7925 | 0.7019 | 8.3249 | 0.7200 | 0.7764 |
+| 11 | LITETime-MV | 11.14 | 0.7506 | 0.7299 | 0.8513 | 0.6820 | 1.3206 | 0.7132 | 0.7637 |
+| 12 | STSF | 11.55 | 0.7724 | 0.7477 | 0.8804 | 0.7080 | 0.6432 | 0.7345 | 0.7826 |
+| 13 | H-InceptionTime | 11.63 | 0.7408 | 0.7190 | 0.8496 | 0.6838 | 1.3227 | 0.7223 | 0.7378 |
+| 14 | LiteTIME | 12.34 | 0.7341 | 0.7104 | 0.8394 | 0.6680 | 1.4776 | 0.7113 | 0.7336 |
+| 15 | PatchMTSC | 13.12 | 0.7428 | 0.6897 | 0.8261 | 0.6533 | 0.7655 | 0.6852 | 0.7352 |
+| 16 | ConvTran | 13.14 | 0.7462 | 0.7102 | 0.8592 | 0.6767 | 0.8190 | 0.7159 | 0.7345 |
+| 17 | Catch22 | 13.15 | 0.7475 | 0.7181 | 0.8697 | 0.6922 | 0.7147 | 0.7240 | 0.7374 |
+| 18 | STC | 13.95 | 0.7545 | 0.7172 | 0.8744 | 0.6940 | 0.6391 | 0.7185 | 0.7537 |
+| 19 | TSF | 14.00 | 0.7515 | 0.7236 | 0.8740 | 0.6883 | 0.7252 | 0.7093 | 0.7606 |
+| 20 | TDE | 14.63 | 0.7262 | 0.6813 | 0.8374 | 0.6382 | 0.8869 | 0.6714 | 0.7344 |
+| 21 | Summary | 16.66 | 0.6858 | 0.6574 | 0.8268 | 0.6230 | 0.9123 | 0.6574 | 0.6844 |
+| 22 | TimesURL | 16.95 | 0.6958 | 0.6533 | 0.7906 | 0.5967 | 1.0055 | 0.6257 | 0.6973 |
+| 23 | 1NN-DTW | 18.45 | 0.6712 | 0.6454 | 0.7197 | 0.6136 | 11.8506 | 0.6521 | 0.6636 |
+| 24 | Dummy | 21.81 | 0.3645 | 0.3029 | 0.5000 | 0.1507 | 1.4067 | 0.2855 | 0.3816 |
+
+Average over the 52 Multiverse-core datasets with results for every estimator on every metric, ordered by average accuracy rank. Best in each column in bold.
 <!-- LEADERBOARD:END -->
 
 Rebuilt with `python -m multiverse.experiments.tables`, which also writes a sortable
@@ -206,7 +213,8 @@ multiverse/
 ├── docs/                    # Documentation
 ├── img/                     # Images used in the documentation
 ├── results/                 # Benchmark results, one directory per classifier
-├── survey/                  # Data behind the MTSC survey
+├── published_results/       # Accuracies reported in the MTSC literature
+├── survey/                  # Signpost only: the directory moved to published_results/
 └── multiverse/              # Python package
     ├── classification/      # Classifiers not available in aeon
     ├── examples/            # Short runnable examples

diff --git a/docs/classifiers.md b/docs/classifiers.md
@@ -28,6 +28,8 @@ from multiverse.classification import (
     PatchMTSCClassifier,
     TimesNetClassifier,
     TimesURLClassifier,
+    TS2VecClassifier,
+    XCMClassifier,
 )
 ```
 
@@ -66,14 +68,61 @@ Wei, Y., et al. "PatchMTSC: patch-based multivariate time series classification"
 
 TimesURL is a self-supervised representation learner rather than an end-to-end
 classifier. A contrastive objective pretrains an encoder on the training collection, the
-collection is encoded, and a logistic regression probe is fitted on those
-representations; at prediction time the fitted encoder embeds the new series and the
-probe classifies them. Ported from the authors'
-[implementation](https://github.com/Alrash/TimesURL).
+collection is encoded, and a probe is fitted on those representations; at prediction time
+the fitted encoder embeds the new series and the probe classifies them. Ported from the
+authors' [implementation](https://github.com/Alrash/TimesURL).
+
+The adapter follows the authors' UEA experiment: channels are standardised using
+training data only, a normalised time coordinate is appended, and an RBF SVM is fitted to
+the full-series representations. `eval_protocol` selects the probe, defaulting to their
+`"svm"`, with `"linear"` and `"knn"` also available.
 
 Liu, J. and Chen, S. "TimesURL: Self-supervised Contrastive Learning for Universal Time
 Series Representation Learning." AAAI, 2024.
 
+## TS2Vec
+
+TS2Vec is a self-supervised representation learner: a hierarchical contrastive objective
+pretrains an encoder, the collection is encoded to one vector per series, and a
+classifier is fitted on the representations. It is the direct comparator to TimesURL,
+which is built on its codebase. Ported from the authors'
+[implementation](https://github.com/zhihanyue/ts2vec).
+
+Yue, Z., Wang, Y., Duan, J., Yang, T., Huang, C., Tong, Y. and Xu, B. "TS2Vec: Towards
+Universal Representation of Time Series." AAAI, 2022.
+
+For UEA the authors evaluate with a support vector machine chosen by grid search over C
+(`train.py` passes `eval_protocol='svm'`), so `probe="svm"` is the default. `probe=
+"logistic"` selects their linear alternative, which is the probe TimesURL uses, so the
+two encoders can be compared without the probe differing between them.
+
+## XCM
+
+XCM is an explainable convolutional network. Two parallel branches see the input
+differently: a 2D branch convolves along time within each channel separately, so its
+activations stay attributable to individual channels, and a 1D branch convolves along
+time across all channels together. The branches are concatenated, passed through a
+further 1D convolution, globally average pooled and classified. The channel
+attributability of the 2D branch is what the paper's explanations rest on, so the
+layer names it refers to are preserved. Ported from the authors'
+[implementation](https://github.com/XAIseries/XCM).
+
+Fauvel, K., Lin, T., Masson, V., Fromont, E. and Termier, A. "XCM: An Explainable
+Convolutional Neural Network for Multivariate Time Series Classification."
+Mathematics, 9(23), 2021.
+
+This is the only Keras port here, following the authors, so it needs `tensorflow`
+rather than `torch`. Both are in the `deep-learning` extra.
+
+The authors tune `window_size` per dataset. Their results table carries a `Win_pct`
+column spread over a five point grid: 20, 40 and 60 on five datasets each, 80 on
+thirteen, and 100 on two. The default here is **0.8**, the value they use most often.
+The 0.2 in their `config.yml` is the worked example for BasicMotions, not a default.
+
+Because `window_size` is a fraction, the kernel grows with the series, and 0.8 of
+EigenWorms' 17984 points would be a 14387 point kernel. `max_window` bounds the kernel
+at 100 points, and it is floored at 1 for very short series.
+
 ## Notes on the ports
 
 All three wrappers take aeon's ``numpy3D`` collections, shape
@@ -91,11 +140,18 @@ package: install it with ``pip install aeon-multiverse[deep-learning]``.
 | ConvTran | [Navidfoumani/ConvTran](https://github.com/Navidfoumani/ConvTran), commit `148afb6` | `Models/model.py`, `Models/Attention.py`, `Models/AbsolutePositionalEncoding.py`, `Models/optimizers.py` |
 | PatchMTSC | [YanxuanWei/PatchMTSC](https://github.com/YanxuanWei/PatchMTSC) | `Models/model.py`, `Models/Attention.py`, `Models/AbsolutePositionalEncoding.py` |
 | TimesURL | [Alrash/TimesURL](https://github.com/Alrash/TimesURL) | the whole model package, vendored under `_timesurl_original` |
-
-ConvTran, PatchMTSC and TimesNet are each a single self-contained module. TimesURL is
-the exception: its implementation spans 11 modules and about 2,300 lines, so it is
-vendored under `multiverse/classification/_timesurl_original` and driven by a thin
-wrapper, rather than inlined. Only the components the published
+| TS2Vec | [zhihanyue/ts2vec](https://github.com/zhihanyue/ts2vec) | `ts2vec.py`, `utils.py`, `models/`, vendored under `_ts2vec_original` |
+| XCM | [XAIseries/XCM](https://github.com/XAIseries/XCM) | `models/xcm.py` |
+
+ConvTran, PatchMTSC and TimesNet are each a single self-contained module. TimesURL and
+TS2Vec are vendored instead, under `_timesurl_original` and `_ts2vec_original`, and
+driven by thin wrappers: they are whole model packages rather than single networks, at
+about 2,300 and 630 lines respectively.
+
+Note that TimesURL is a fork of TS2Vec, not a user of it. Its copies of `encoder.py` and
+`losses.py` are modified: the loss calls mixup variants that TS2Vec does not have, and
+the encoder carries a reconstruction head. The two vendored packages are therefore kept
+separate rather than sharing a base. Only the components the published
 architecture actually reaches are reproduced: for ConvTran and PatchMTSC that is the
 ``tAPE`` fixed encoding and ``eRPE`` relative encoding path, so the alternative
 encodings (``Sin``, ``Learn``, ``Vector``) and the unused ``Transformer`` and
@@ -120,7 +176,10 @@ faithful transcription.
 | Sibling imports rewritten as relative imports | TimesURL | Upstream is laid out for ``sys.path`` insertion; as a subpackage it needs relative imports. A stray package-level ``from .encoder import TSEncoder`` is dropped, since ``encoder`` lives under ``models`` and the import only went unnoticed because the ``sys.path`` route bypassed it |
 | One unconditional ``print`` silenced | TimesURL | ``lib.py`` printed the training tensor shape on every fit, which would pollute benchmark logs. ``verbose`` controls training output instead |
 | Python's ``random`` seeded alongside numpy and torch | TimesURL | The authors' collator draws from ``random`` for segment masking and index shuffling, so seeding numpy and torch alone left runs irreproducible |
-| ``multi_class`` dropped from the probe | TimesURL | The authors pass ``multi_class="auto"``, deprecated in scikit-learn 1.5 and removed in 1.8. The default already matches its behaviour here |
+| Probe probabilities from ``decision_function`` | TimesURL | An ``SVC`` without ``probability=True`` cannot produce probability estimates, and aeon classifiers must implement ``predict_proba``. The decision scores are softmaxed instead, which avoids the internal cross-validation that Platt scaling would add |
+| ``probability=True`` on the SVM probe | TS2Vec | The authors' grid sets it False, which leaves an ``SVC`` unable to produce probability estimates. aeon classifiers must implement ``predict_proba``, so it is enabled, adding Platt scaling fitted by internal cross-validation on the training data |
+| Layer imports taken from ``tensorflow.keras.layers`` | XCM | The original imports ``Conv1D`` and ``Conv2D`` from ``keras.layers.convolutional``, a path removed in Keras 3. The layers and their arguments are unchanged |
+| Kernel length floored at one point | XCM | The original computes ``int(window_size * n)``, which is zero for series shorter than five points and builds an invalid layer |
 | Validation split moved inside ``fit`` | all three | The originals split train/validation outside the model, which risks leakage between train and test. TSLib is explicit about it: ``exp_classification.py`` sets ``vali_data = self._get_data(flag='TEST')``, so it selects the retained epoch on the test set. See the note at the top of this page |
 | Test data scaled with training statistics | TimesNet | TSLib fits its normaliser separately per split, so its test set is scaled by its own statistics. The port fits on train and applies to test |
 
@@ -167,6 +226,8 @@ tests record which ports have which:
 |---|---|---|
 | ConvTran | yes | yes |
 | TimesURL | vendored verbatim, so identical by construction | n/a |
+| TS2Vec | vendored verbatim, so identical by construction | n/a |
+| XCM | yes, the built graph matches layer for layer with equal parameter counts | n/a, Keras rather than torch |
 | PatchMTSC | yes | no, the original's head is an ``nn.LazyLinear`` that draws its weights on the first forward pass rather than at construction |
 | TimesNet | yes | no, TSLib always builds a temporal embedding that the classification path never applies; dropping it removes one weight tensor and shifts every later draw |
 
@@ -182,6 +243,8 @@ export MULTIVERSE_CONVTRAN_SRC=/path/to/ConvTran
 export MULTIVERSE_PATCHMTSC_SRC=/path/to/PatchMTSC
 export MULTIVERSE_TIMESNET_SRC=/path/to/Time-Series-Library
 export MULTIVERSE_TIMESURL_SRC=/path/to/TimesURL/model/modules
+export MULTIVERSE_TS2VEC_SRC=/path/to/ts2vec
+export MULTIVERSE_XCM_SRC=/path/to/XCM
 pytest multiverse/classification/tests/test_original_equivalence.py -v
 ```
 

diff --git a/multiverse/classification/__init__.py b/multiverse/classification/__init__.py
@@ -7,10 +7,14 @@
     "ConvTranClassifier",
     "PatchMTSCClassifier",
     "TimesNetClassifier",
+    "TS2VecClassifier",
+    "XCMClassifier",
     "TimesURLClassifier",
 ]
 
 from multiverse.classification._convtran import ConvTranClassifier
 from multiverse.classification._patchmtsc import PatchMTSCClassifier
 from multiverse.classification._timesnet import TimesNetClassifier
+from multiverse.classification._ts2vec import TS2VecClassifier
+from multiverse.classification._xcm import XCMClassifier
 from multiverse.classification._timesurl import TimesURLClassifier
diff --git a/multiverse/classification/_timesnet.py b/multiverse/classification/_timesnet.py
@@ -92,13 +92,22 @@ class TimesNetClassifier(BaseClassifier):
           Maximum number of training epochs.
       learning_rate : float, default=1e-3
           Learning rate for RAdam.
-      lr_adjust : {"type1", "cosine", None}, default="type1"
+      lr_adjust : {"type1", "cosine", None}, default=None
           Learning rate schedule, applied every five epochs as in the original
-          TSLib training loop. ``"type1"``, the TSLib default, sets the rate to
-          ``learning_rate * 0.5 ** (epoch - 1)`` at epochs 5, 10, 15, and so on,
-          which decays it to near zero part way through a default 30 epoch run.
-          ``"cosine"`` follows TSLib's cosine option. ``None`` disables the
-          schedule and trains at a constant rate.
+          TSLib training loop. ``"type1"`` is TSLib's own default and sets the
+          rate to ``learning_rate * 0.5 ** (epoch - 1)`` at epochs 5, 10, 15 and
+          so on, which from the published ``learning_rate=0.001`` reaches 2.0e-6
+          by epoch 10 and 6.1e-8 by epoch 15, so training effectively stops
+          about a third of the way through a 30 epoch run.
+
+          The default here is None, which departs from TSLib deliberately. Their
+          schedule is harmless in their pipeline because they select the
+          retained epoch on the *test* set, so an early epoch from before the
+          collapse is kept anyway. This wrapper selects on a held-out split of
+          the training data, so it keeps a model that stopped learning. On ERing
+          the difference is 0.578 with the schedule against 0.933 without, where
+          the published result is 0.915. Set ``"type1"`` to reproduce TSLib's
+          behaviour, or ``"cosine"`` for their cosine option.
       patience : int, default=10
           Early stopping patience based on internal validation accuracy.
       validation_size : float, default=0.2
@@ -178,7 +187,7 @@ def __init__(
         batch_size: int = 16,
         n_epochs: int = 30,
         learning_rate: float = 1e-3,
-        lr_adjust: str | None = "type1",
+        lr_adjust: str | None = None,
         patience: int = 10,
         validation_size: float = 0.2,
         gradient_clip: float = 4.0,