Changelog

All notable changes to this project will be documented in this file.

[0.3.0] - 2025-01-31

Added

  • Implemented ClusterBasedLocalOutlierFactor (CBLOF) anomaly detector.

  • Implemented KMeansAnomalyDetector anomaly detector.

  • Implemented CopulaBasedOutlierDetector (COPOD) anomaly detector.

  • Implemented RobustScaler preprocessor.

  • Added optional dependencies to dtaianomaly, rather than having to install everything via different requirements files.

  • Added support for Python 3.13.

  • Implemented the predict_confidence() method in the BaseDetector, which computes a confidence score for each prediction of the anomaly detector.

  • Integrated black and isort to format the code.

Changed

  • Added kwargs parameters to data.read_from_directory() to pass additional arguments to data loaders.

  • Added a check for BinaryMetric if the metric is actually binary or not.

  • Separated the resource requirements in the ´´Workflow`` across the fit and predict stage of the anomaly detector. The aggregated resources are still returned as before. For this, the EvaluationPipeline received several new functions which execute each stage in an atomic manner.

  • When failing to automatically compute a window size, the default behavior is to raise a ValueError (compared to returning -1 before). This makes it clear where the something goes wrong, instead of propagating the faulty window size to for example the sliding_window() function. Additionally, it is possible to set a default window size in compute_window_size(), which will be returned (if provided) instead of raising an error to allow the system to continue.

Fixed

  • Renamed ZNormalizer to StandardScaler, to make it align with the Sklearn declaration.

[0.2.3] - 2024-12-02

Added

  • Added option to automatically compute the window size for various anomaly detectors using Fourier Transform, Autocorrelation, Multi-Window Finder, or Summary Statistics Subsequence.

  • Implemented KNearestNeighbor anomaly detector.

  • Implemented point-adjusted precision, recall and FBeta.

  • Implemented BestThresholdMetric, a ProbaMetric which computes the best value for a BinaryMetric over all thresholds.

  • The property supervision to BaseDetector, which indicates what type of supervision the anomaly detector requires. Possible options are:

    • Supervision.SUPERVISED: the anomaly detector requires training data and labels

    • Supervision.SEMI_SUPERVISED: the anomaly detector requires training data, but no training labels because the training data is assumed to be normal.

    • Supervision.UNSUPERVISED: the anomaly detector does not require any training data or labels

  • Added the property __version__ to dtaianomaly, which can be accessed from code.

  • Included the used version of dtaianomaly when logging errors.

  • Added the plot_time_series_anomalies function to visualize time series data with real and predicted anomalies, highlighting TP, FP, and FN cases.

  • Implemented PrincipalComponentAnalysis, KernelPrincipalComponentAnalysis and RobustPrincipalComponentAnalysis anomaly detectors.

  • Implemented HistogramBasedOutlierScore anomaly detector.

  • Implemented OneClassSupportVectorMachine anomaly detector.

  • Implemented methods to visualize time series and anomalies: plot_demarcated_anomalies, plot_with_zoom, and plot_anomaly_scores.

Changed

  • Updated documentation to contain changelog and contributing information

  • Rely on PyOD for non-time series anomaly detectors (instead of scikit-learn before)

  • Separated training data and test data in DataSet. This has also been integrated within the Workflow to use the correct data. To this end, a new flag has been added to the Workflow, which decides if the training data or the test data should be used for training unsupervised anomaly detectors.

  • Detectors implementing PyODAnomalyDetector must implement an additional method to declare the Supervision of that detector.

Fixed

[0.2.2] - 2024-10-30

Added

  • When executing a workflow, and an error occurs. The errors will be written to an error file. This file contains in which phase the error occurred and the entire traceback of the error. Additionally, the error file contains the code to reproduce the specific error. In fact, the error file can be run as any python script.

  • Added baseline anomaly detectors: AlwaysNormal, AlwaysAnomaly, and RandomDetector

  • Added option novelty to MatrixProfileDetector, which will compute the matrix profile in regard to the train data, if novelty=True. By default, the matrix profile is computed based on a self-join of the test data.

  • Implemented MedianMethod anomaly detector.

  • Implemented Differencing preprocessor.

  • Implemented PiecewiseAggregateApproximation preprocessor.

Changed

  • Added the option to cache data in LazyDataLoader via parameter do_caching. The load function in LazyDataLoader is adjusted to either load the data and potentially cache the data, or return a cached version of the data. As a consequence, the children of LazyDataLoader must implement the _load method (instead of the load() method), which will effectively load the data, independent of any cache.

Fixed

  • utils.is_valid_array_like() could not handle multivariate lists. This functionality has now been added, and the tests are extended accordingly.

  • Preprocessors can now take lists as input, which are automatically converted to a numpy array by the fit() and transform() method in Preprocessor.

[0.2.1] - 2024-10-08

In this release, all files were transferred from GitLab to GitHub. Therefore, older links may no longer function as desired.

Added

Changed

  • The __str__ method of the different estimators are automatically done and now equal the name of the class and between parantheses the parameters that must be provided, i.e., the parameters that differntiate from the default parameters.

Fixed

[0.2.0] - 2024-10-01

This release basically revamped the dtaianomaly. In many ways, the package has been simplified, while also ensuring its robustness. In general, the core functionality remained similar, while the API might have slightly changed. Below we mention the most notable changes.

Added

  • A LazyDataLoader has been added, which can be used to read datasets from disk. This offers a simpler interface compared to the old DataManager.

  • A Pipeline has been added to easily combine time series anomaly detection with preprocessing the time series

  • The Preprocessing module has been added, which includes a wide range of methods to preprocess a time series before detecting anomalies. Specifically, the implemented preprocessors are ExponentialMovingAverage, MinMaxScaler, MovingAverage, SamplingRateUnderSampler, NbSamplesUnderSampler, and Znormalizer. In addition, there is also a ChainedPreprocessor which combines multiple preprocessors.

  • A Workflow object, which takes a set of dataloader, anomaly detectors, preprocessors and evaluation metrics and evaluates in a grid-like manner each anomaly detector in combination with each preprocessor on each dataset. As before, a workflow can be instantiated from a configuration file, but now it is also possible to start a workflow from Python itself, meaning that the Python scripts can serve as configuration files.

  • More tests have been added to ensure dtaianomaly functions well and to guarantee that changes do not alter the existing functionality.

Changed

  • The base anomaly detector has been renamed from TimeSeriesAnomalyDetector to BaseDetector. Additionally, the wrapper approaches to existing libraries for anomaly detection have been removed, as these rely on the active maintenance of said packages (specifically, PyODAnomalyDetector and TSBUADAnomalyDetecotor have been removed).

  • The evaluation module has been cleaned up to only contain well-established evaluation metrics. In the future, we plan on adding more performance metrics.

  • The thresholding methods have been moved from the evaluation module into their own module: thresholding.

  • The notebooks have been cleaned up to only show the core functionality to the users, making it easier to understand dtaianomaly.

Fixed

  • The visualization methods were relatively buggy. Most of the methods have been removed from this module, because it is simple to plot time series and the anomaly scores using plt.plot(X) and plt.plot(y). Only one method remained to plot a time series and color it according to the anomaly scores, as this is not trivial to do with just matplotlib.

  • The documentation has improved, including the API descriptions, but especially the getting started pages.

[0.1.4] - 2024-02-28

This release mostly increased the amount of functionality, but also offers some more quality-of-life features.

Added

  • TSB-UAD has been integrated, thus increasing the amount of available algorithms.

  • Options to read the results from a workflow and visualize them.

  • An option has been added to log errors in the workflow, without letting the entire workflow crash and stop.

  • Anomaly detector STOMP (based on the Matrix Profile) has been added.

  • An option to include a specific stride when windowing the time series using the Windowing class has been added.

Changed

  • Due to the dependency on TSB-UAD, this version (and likely also future versions) won’t be available on PyPi anymore, because TSB-UAD is installed from source, which means is not supported through PyPi

  • Changed how the algorithm configuration works.

    • you can provide multiple algorithms in one configuration to facilitate large scale experiments in which multiple algorithms are compared.

    • An option was implemented to provide template configurations, and then fill in the templates given a number of possible values in a grid-like fashion. this allows to more easily tune various parameters of anomaly detectors.

  • The number of features in the DataManager are reduced such that only a limitted set of important features remain.

Fixed

  • Some bugs related to visualizing the data have been fixed.

  • There was a problem with using custom algorithms in the workflow, due to an unknown path.

  • Added the opportunity to perform anomaly detection in parallel over multiple time series, thus reducing the total required running time.

[0.1.3] - 2023-11-07

There was another, similar bug.

Added

Changed

Fixed

  • Also added a __init__.py file in the utility directory for the affiliation metrics.

[0.1.2] - 2023-11-07

This update is to fix a crucial but small bug.

Added

  • The documentation has been extended (though far from finalized).

Changed

Fixed

  • The __init__.py files in the anomaly_detection module were updated to properly import classes that are not directly in the anomaly_detection, but rather in a sub folder.

[0.1.1] - 2023-10-26

This update doesn’t include a lot of changes. It only slightly modified the readme.

Added

  • Added an official release to the repository, and a badge to indicate the latest release.

Changed

Fixed

  • Fixed the link to the image showcasing the anomaly scores of an IForest on a Demo time series.

[0.1.0] - 2023-10-26

First release of dtaianomaly! While our toolbox is still a work in progress, we believe it is already in a usable stage. Additionally, by publicly releasing dtaianomaly, we hope to receive feedback from the community! Be sure to check out the [documentation](https://u0143709.pages.gitlab.kuleuven.be/dtaianomaly/) for additional information!

Added

  • anomaly_detection: a module for time series anomaly detection algorithms.

    Currently, basic algorithms using[PyOD](https://github.com/yzhao062/pyod) are included, but we plan to extend on this in the future!

  • data_management: a module to easily handle datasets. You can filter the datasets on

    certain properties and add new datasets through a few simple function calls! More information can be found in the [Documentation](https://u0143709.pages.gitlab.kuleuven.be/dtaianomaly/getting_started/data_management.html).

  • evaluation: It is crucial to evaluate an anomaly detector in order to quantify its

    performance. This module offers several metrics to this end. dtaianomaly offers traditional metrics such as precision, recall, and F1-score, but also more recent metrics that were tailored for time series anomaly detection such as the [Affiliation Score](https://dl.acm.org/doi/10.1145/3534678.3539339) [notebooks](notebooks) and [Volume under the surface (VUS)](https://dl.acm.org/doi/10.14778/3551793.3551830)

  • visualization: This module allows to easily visualize the data and anomalies, as

    time series and anomalies inherently are great for visual inspection.

  • workflow: This module allows to benchmark an algorithm on a larger set of datasets,

    through configuration files. This methodology ensures reproducibility by simply providing the configuration files!

Changed

Fixed