Changelog

This changelog highlights notable updates. For full commit history, refer to GitHub.

Unreleased

• Add a CITATION.cff file, read by GitHub’s Cite this repository button, and the citation of the TSB-SEG benchmark (NeurIPS 2026) to the README.

0.3.0 (2026-10-14)

Upgrading from 0.2.0: VQTSSDetector, VSAXDetector and the HDP-HSMM aliases are removed; GreedyGaussianDetector returns change points instead of labels; the arguments of SNLDSDetector are keyword-only; ProphetDetector, and ChangeFinderDetector without n_cps, select change points differently by default; BeastDetector needs Rbeast from PyPI (tsseg[beast]); the supervision and dependency tags of several detectors change. KCPDDetector, HMMDetector, PatssDetector and the n_cps parameter of AutoPlaitDetector are deprecated. Several fixes change the output of AutoPlait, BEAST, IGTS, Window and the vendored ruptures solvers (ties): results computed with 0.2.0 can differ.

• New detector ChangeForestDetector: classifier-based nonparametric change point detection with random forests (Londschien, Bühlmann and Kovács, JMLR 2023). It wraps the official changeforest package, installed with the new extra pip install tsseg[changeforest]. With n_cps it runs best-first binary segmentation on the package’s gains, a semi-supervised mode that is not part of the paper.

• New RuLSIFDetector (Liu, Yamada, Collier and Sugiyama, Neural Networks 2013): change points are peaks of the symmetrised relative Pearson divergence between the subsequences before and after each point, estimated by RuLSIF with cross-validated kernel width and regularisation. NumPy port of the authors’ MATLAB code (MIT); guided (n_cps) or thresholded.

• TSCP2Detector takes a random_state seed (None by default, as before): with an integer, every fit initialises the encoder, shuffles the training windows and draws the dropout masks identically, so two fits on the same series return the same change points. Without a seed, the weights and the shuffling changed at every fit, and nothing declared the detector non-deterministic. The seed goes through tf.keras.utils.set_random_seed, which also reseeds the global Python, NumPy and TensorFlow generators.

• Dependencies: tqdm and joblib are no longer required. No module of tsseg imports them since the unfinished Python port of AutoPlait was removed (joblib remains installed with scikit-learn). fetch_dataset shows a download progress bar only when tqdm is installed.

• Tags: the undocumented semi_supervised tag of AmocDetector, EspressoDetector and RandomDetector is removed: capability:semi_supervised is the tag (the semi_supervised parameter of RandomDetector is unchanged). The deprecated HMMDetector, neither unsupervised nor semi-supervised, declares capability:semi_supervised: its number of states is the length of emission_funcs. A contract test rejects unknown tags and checks that every detector declares both supervision tags, at least one of them True.

• Tags: python_dependencies is None or a list of the optional modules a detector imports. ProphetDetector, TGLADDetector and TireDetector declared a string ("torch,networkx"), and eight detectors declared nothing; they now declare ["numba", "aeon"] (ClapDetector), ["numba"] (ClaspDetector, EAggloDetector), ["torch"] (E2USDDetector, Time2StateDetector), ["stumpy"] (FLUSSDetector), ["npbad"] (PatssDetector) and ["Rbeast"] (BeastDetector). A contract test checks the format and that the tag lists the modules the test suite skips a detector on.

• make test and make docs run the test and documentation jobs of the CI (pytest --tb=short -q; sphinx -W -E -b html docs docs/_build/html). PYTEST_ARGS selects tests, DOCS_BUILD_DIR changes the output directory, PYTHON the interpreter.

• Reproduction tests: tests that reproduce a published result on a dataset downloaded on demand (pinned by its SHA-256 digest) go to tests/reproduction/ and are marked reproduction. The marker is deselected by default (pytest -m reproduction runs them); the protocol is in tests/reproduction/README.md.

• GreedyGaussianDetector returns the change point indices, as its returns_dense=True tag and its documentation announce: it returned one segment label per time point (0, ..., 0, 1, ..., 1, ...) since 0.2.0, which code reading the output as change point indices misinterpreted.

• EAggloDetector returns a numpy integer array instead of a Python list, and the deprecated HMMDetector returns integer labels instead of floats.

• InformationGainDetector stops its search when every candidate position (the multiples of step) is already a change point, and then returns fewer than k_max change points: when k_max exceeded the number of candidates (ceil(n / step) - 1), it returned -1 and 0 among the change points, for instance [-1, 0, 5, 10, 15, 20, 25] for k_max=10, step=5 on 30 points. The other change points are unchanged.

• Code that no module of the package imports is removed: an unfinished Python port of AutoPlait, older BOCD and TIRE scripts, the former IGTS implementation, a MATLAB-engine wrapper of ESPRESSO, PaTSS scripts that import a baselines package foreign to tsseg, and the MATLAB reference sources of ESPRESSO and iCID. The wheel no longer ships these modules nor the four data files of the MATLAB sources (two .mat files of the iCID demo, two .txt notes of ESPRESSO). The deprecated aliases HdpHsmmDetectorV1, HdpHsmmLegacyDetector and LegacyHdpHsmmDetector of tsseg.algorithms.hdp_hsmm are removed: use HdpHsmmDetector.

• AutoPlaitDetector returns the segmentation that the AutoPlait binary writes: each time point gets the number of its regime, so that segments of the same regime share a label, and change points are the first index of a segment. The wrapper read the change points one index early (the last index of the previous segment), took the labels of segment.labels (one row per regime, whose second column is the regime’s path in the split tree) and applied them to the segments in time order, padding with the last one: on regimes A-B-A it returned 1 | 0 | 0 with a single change point. It now raises RuntimeError when the binary’s segments do not cover each time point exactly once.

• AutoPlaitDetector is tagged unsupervised (capability:unsupervised True, capability:semi_supervised False), as the AutoPlait binary always ran: it chooses the number of segments itself and has no guided mode. It runs without n_cps and without y; it raised ValueError when both were missing. n_cps is deprecated and ignored: passing it, 0 included, emits a FutureWarning at fit and no longer launches the binary once with a number of segments that it rejected. The output does not change.

• AutoPlaitDetector hands the AutoPlait binary a series that it reads with the right length. The binary counts the time points of its input with a line buffer of 254 bytes, and the wrapper wrote 25 characters per value (%.18e): rows of 10 channels or more were counted twice or more, and the binary segmented a longer series padded with rows of zeros (62,442 time points instead of 31,392 on ActRecTut subject 1), whose last segment ended past the series. Values are now written as float32 with 9 significant digits, the precision at which the binary reads them, so that series of up to 15 channels always fit. A series whose rows still do not fit (the 76 channels of JIGSAWS) raises ValueError before the binary runs, and a binary that reports another length than the series’ raises RuntimeError. Series of fewer than 10 channels were not affected.

• The AutoPlait runner (AutoPlait._run_autoplait in tsseg.algorithms.autoplait.autoplait_c) runs the binary once, with the command line of the bundled binary (autoplait d list_file out_dir), and raises RuntimeError if it fails. It no longer takes n_cps, nor tries a command line with n_cps + 1 or with the length of the series: the only binary that takes a second argument, the modified version of the FLOSS project page, reads it as a minimum number of regimes, not of segments. The documentation now points to the original C code of Matsubara et al., which c/autoplait/ holds, instead of the FLOSS page. AutoPlaitDetector returns the same output.

• The output contract tests check what the documentation states: change points are a 1-D integer numpy array, strictly increasing, within [1, n - 1]; state labels are a 1-D integer numpy array of length n. They accepted 0, duplicates and float labels.

• Guided mode on a series without any change point: ClaspDetector (n_change_points=0 or n_segments=1), ClapDetector (n_change_points=0), EspressoDetector (n_segments=1), Time2StateDetector, E2USDDetector and TiccDetector (n_states=1) accept the value and return no change point (a single state); their parameter validation rejected it.

• AmocDetector computes the error of every split from cumulative sums, in linear time: it recomputed both means at every split (quadratic time). The detected change point is the same.

• fit and fit_predict reject a y given as a pandas.DataFrame that does not have exactly one column, as the error message states: a frame with two columns (or none) was accepted.

• ChangeFinderDetector without n_cps: the data-driven threshold is mean + threshold_factor * std of the scores (threshold_factor, 2.0 by default as before), the minimum distance between change points can be a fraction of the length (min_distance_fraction), and the ensembling strategy merges the detections of the channels that lie within the tolerance (by default the minimum distance) and can require a share of the channels to agree (consensus, 0 by default); it returned every distinct peak of every channel. The l2 strategy and the behaviour with n_cps are unchanged by default.

• FLUSSDetector accepts n_segments=None (number of segments unknown): the regime extraction keeps the minima of the corrected arc curve below threshold (0.5), and the ensembling strategy merges the detections of the channels within tolerance (consensus, the share of the channels that must agree, is 0 by default). A known n_segments behaves as before.

• New helper tsseg.algorithms.utils.consensus_change_points: change points detected by enough channels, without a target count.

• Documentation matches the code: the tag list of the detectors page (returns_dense is True for change point indices; guided mode is the capability:semi_supervised tag), the getting-started guide (examples that run without optional dependencies, guided mode through constructor parameters, real test commands), the datasets page (one bundled dataset, no test fixture to import), the extras table, change point examples that named their output labels, and the citation (Springer proceedings).

• Docstrings: PeltDetector, BottomUpDetector, WindowDetector, DynpDetector and KCPDDetector document their parameters; BinSegDetector states its actual penalty default (10; pass penalty=None to stop on epsilon); the HidalgoDetector example runs with tsseg instead of aeon, and its K_states default is 1; the ICIDDetector psi_list default is None. The compute methods of the state detection metrics and the validate methods of the parameter constraints are documented. A test checks that each detector documents exactly its constructor parameters, with their actual defaults.

• EAggloDetector: penalty="mean_diff_penalty" no longer raises (the penalty was compiled with numba, which rejects the list of segment boundaries it receives). The documentation now describes the goodness-of-fit statistic, the V-statistics and the merge of the first and last clusters inherited from ecp, and the over-segmentation obtained from one cluster per time point.

• ClaspDetector runs when NUMBA_NUM_THREADS is below the number of cores (as set by users or by SLURM jobs): n_jobs=-1 meant os.cpu_count() threads, more than numba accepts, which raised ValueError: The number of threads must be between 1 and .... n_jobs=-1 now means all the threads numba may use, and a larger n_jobs is capped at that number.

• TireDetector: the minimum distance between two change points is peak_distance_fraction times the series length, as documented; the default is 0 (no constraint). The prominence threshold of the unsupervised mode is the prominence_threshold parameter, 0.5 by default.

• Removed VQTSSDetector and VSAXDetector, two methods developed for tsseg without a publication. On a 54-series sample of TSB-SEG, VSAX scored below a random segmentation (bidirectional covering 0.351 against 0.377, ARI 0.163 against 0.208), and VQTSS returned one vector-quantisation code per time step rather than a segmentation (ARI 0.001), with more than 10 minutes per series beyond 2,500 points. Both labelled local patterns rather than regimes. Code that imports them must drop them; they remain in the history of the repository and in tsseg 0.2.0.

• ProphetDetector now places n_candidates evenly spaced candidates, fits Prophet’s rate changes delta and returns the n_changepoints largest |delta| relative to their median, or, when n_changepoints is None (the new default), those above delta_threshold (a relative form of TCPDBench’s rule), at least tolerance apart. New parameters n_candidates, changepoint_prior_scale (500 by default, calibrated on TSB-SEG), delta_threshold, seasonality (off by default) and max_points (longer series are reduced to block means before the fit); n_changepoints now accepts 0; multivariate_strategy defaults to "l2", and "ensembling" averages the per-channel scores instead of voting.

• The vendored ruptures solvers have a numba backend for the l1, l2, rbf and cosine costs, with the same segmentation as the Python path: backend="auto" (the default) picks it when numba is installed, "numba" and "python" force either path, in PeltDetector, DynpDetector, BinSegDetector, BottomUpDetector, WindowDetector and KCPDDetector. The kernel costs no longer build the n x n Gram matrix (for PELT and DynP, a transcription of the C solvers of ruptures’ KernelCPD); DynP no longer memoises O((n / jump)^2) sub-problems; BinSeg finds the best split of a segment in one sweep from each end.

• SNLDSDetector is now shipped with the package. It was listed in tsseg.algorithms without its module, so from tsseg.algorithms import * failed on the 0.2.0 wheel. The authors’ TensorFlow code (Google Research, Apache-2.0) is vendored in tsseg/algorithms/snlds/vendor; install the dependencies with pip install tsseg[snlds]. Its constructor is now keyword-only, and its tags declare it semi-supervised (n_states is the number of regimes) rather than unsupervised. The wheel now also ships the LICENSE files of the vendored code.

• CostRbf: the median heuristic no longer counts the distance of a sample to itself, whose rounding residues shifted gamma by about 1e-4, and no longer builds the Gram matrix.

• Ties: every vendored solver treats values closer than 1e-9 of the costs involved (at least of the signal’s cost per sample, tie_unit) as equal, and decides between them by position as upstream decides between exactly equal values: earliest change point (PELT, DynP), last change point of a segment and first segment (BinSeg), leftmost merge (BottomUp), no peak on a plateau and latest tied peak first (Window); a value tied with pen or epsilon stops, and PELT keeps a start tied with its pruning bound (pruning less never loses the optimum).

• PatssDetector is deprecated (FutureWarning) and will be removed in a future release: it needs npbad, which is not on PyPI, and over-segments (17 times more change points than the ground truth on a 54-series sample of TSB-SEG). The authors maintain PaTSEmb, which needs Java.

• WindowDetector searches its peaks as ruptures does (argrelmax(mode="wrap")): the vendored peak search cut the neighbourhood of a score at the ends of the signal, and could add a change point next to the first or last window.

• The vendored ruptures costs convert the signal to float64 in fit: float32 signals were costed in float32, and CostCosine failed on integers.

• BeastDetector uses the Rbeast package from PyPI, version 0.1.25 or later (new extra tsseg[beast]), instead of the copy of Rbeast 0.1.15 vendored in c/Rbeast, which never returned on univariate series longer than about 4,600 points. Change points were one sample late (Rbeast counts time from 1); they are now the index of the first sample of the new segment. mcmc_seed defaults to 1: with 0, Rbeast draws a random seed and the output changed from one call to the next.

• Third-party licenses: bocd/, ticc/, time2state/ and the aeon-derived eagglo/, hidalgo/, hmm/ and igts/ now ship a copy of their upstream license, in the source tree and the wheel. The code adapted in bocd/ is under the MIT license, not Apache-2.0 as documented. The README lists BeastDetector’s dependency Rbeast, under the license of its R package (GPL >= 2), and no longer claims that every vendored directory contains a license file.

• pen_scale="bic" in PeltDetector, BinSegDetector, BottomUpDetector and WindowDetector: the penalty is a coefficient on log(n) * u, u the cost of one sample of the whole signal (d for l2 on unit-variance channels, where it is the BIC penalty; the mean absolute deviation for l1; 1 for the kernel costs), so that one value applies across lengths, dimensions and costs.

• New GaussianHMMDetector: a hidden Markov model with diagonal Gaussian emissions, fitted by Baum-Welch (EM) from k-means initialisations and decoded by Viterbi. The number of states is given (n_states) or chosen by BIC. HMMDetector, which only decoded with user-supplied parameters, is deprecated (FutureWarning).

• BinSegDetector, BottomUpDetector and WindowDetector accept n_cps=0 again (no change point), as DynpDetector and KCPDDetector do: the parameter schema required n_cps >= 1, which failed the semi-supervised runs on series without change points.

• SNLDSDetector follows the training schedules of the paper: the temperature of the discrete transitions and the weight of the cross-entropy regulariser start at 1,000 (temperature_init, xent_init) and decay as in the Reacher experiment, scaled to n_train_steps. Before, both stayed inactive (temperature 1, weight 1e-10). Training draws batch_size windows of train_window points per step (new parameter, default 256; None for the whole series), so a step no longer runs over the whole series, and the defaults are batch_size=8 and n_train_steps=2000. Prediction is compiled, about 100 times faster on MoCap. The detector declares a parameter schema and the non_deterministic tag, and no longer reseeds NumPy’s global generator. SNLDS is documented as experimental: on a sample of TSB-SEG it scores below a random segmentation.

• KCPDDetector is deprecated (FutureWarning): it is PeltDetector or DynpDetector with a kernel cost (model="rbf", "cosine", or "l2" for the linear kernel).

• PatssDetector runs with scikit-learn 1.8, which removed LogisticRegression(multi_class=...) and raised a TypeError: the one-vs-rest elastic-net logistic regression of PaTSS is now OneVsRestClassifier(LogisticRegression(...)), the same model (predicted probabilities equal to 1e-15 under scikit-learn 1.6).

• F1Score gives 1.0 when neither the ground truth nor the prediction has a change point and the series length is read from y_true: F1Score().compute([n], []) gave 0.0, because the empty prediction was rejected before the boundaries were set aside (compute([n], [n]) and compute([n], [], n_timepoints=n) already gave 1.0). Scores computed from labels (convert_labels_to_segments=True) are unchanged. The warning about a margin rounded to 0 samples is only emitted when change points are matched.

• F1Score, Covering, BidirectionalCovering and GaussianF1Score accept the series length in compute(y_true, y_pred, n_timepoints=n); the change points may then include or omit 0 and n, so the raw output of a detector can be scored directly. Without n_timepoints (and without convert_labels_to_segments=True), the length is still read from the change points, scores are unchanged, and a UserWarning now says so: a y_true that does not end with the series length silently lost its last change point (F1Score(margin=5).compute([300, 700], [300]) gives 1.0; with n_timepoints=1000, 0.667).

• Detector pages give the time complexity of 18 more detectors, as implemented with their default settings. Two values were wrong: PELT is \(O(C\,n)\) expected and \(O(C\,n^{2})\) at worst, with no factor in the number of change points (the page said \(O(C\,K\,n)\)), and ChangeFinder solves one \(k \times k\) system per sample, so \(O(d\,n + n\,k^{3})\) (the page said \(O(n\,k^{2})\)).

• The tGLAD page gives the default of epochs (100; it said 2000) and explains why the paper reports \(O(n)\) where the page states \(O(d^{3}\,n)\) per epoch. The TGLADDetector docstring links to the authors’ repository (Harshs27/tGLAD).

0.2.0 (2026-09-26)

• Replaced the NumPy HdpHsmmDetector with a corrected sampler equivalent to pyhsmm’s WeakLimitHDPHSMM (auxiliary self-transitions of Johnson & Willsky, right-censored last segment, chain started from the prior, numba forward messages) and calibrated defaults.

• numba is optional again, as tsseg[accelerators] announces: ClapDetector, ClaspDetector and EAggloDetector are imported lazily (they need numba), and HdpHsmmDetector falls back to interpreted code without it.

• EAggloDetector and GreedyGaussianDetector declare returns_dense=True, consistent with their change point output.

• KCPDDetector: decimation parameter, pen_scale="bic" (penalty as a coefficient on log(n) * d), and n_cps=0 accepted.

0.1.1 (2026-04-20)

• Added BeastDetector and ChangeFinderDetector.

• InformationGainDetector (IGTS) extended to univariate series, with the optimisation of the paper.

• Parameter schemas for all detectors; GF1 score exposed.

• Fixes to GLAD, Hidalgo and VSAX; continuous integration (lint, tests, docs).

0.1.0 (2026-02-10)

Initial public release.

• Vendored a lightweight subset of ruptures v1.1.8 — BinSeg, BottomUp, DynP, KCPD, PELT and Window detectors no longer need an external dependency.

• Added BottomUpDetector and DynpDetector wrappers with aeon-compatible interfaces.

• Replaced the surrogate PyTorch TSCP2Detector with a faithful TensorFlow + TCN implementation matching the original baseline.

• Replaced the pyhsmm-backed HdpHsmmDetector with a native NumPy/SciPy sticky HDP implementation. pyhsmm and pybasicbayes are no longer installed by default.

• Improved ClaspDetector axis handling.

• Added Sphinx documentation with GitHub Pages deployment.

• Documented test suite layout and fixtures.