Changelog¶

0.5.0, 21 April 2024¶

  • This release makes more changes to the tests and documentation in reponse to the review process for the submission to JOSS (see below).

  • In particular, see the following issue: #97

  • Changed the method of handling dynamic versioning. For now the package __version__ attribute is still defined, but it is deprecated and will be removed in 0.6.0. Use from importlib.metadata.version('redflag') to get the version information instead.

  • Changed the default get_outliers() method from isolation forest ('iso') to Mahalanobis ('mah') to match other functions, eg has_outliers() and the sklearn pipeline object.

  • Updated actions/setup-python to use v5.

0.4.2, 10 December 2023¶

  • This is a minor release making changes to the tests and documentation in reponse to the review process for a submission to The Journal of Open Source Software (JOSS).

  • See the following issues: #89, #90, #91, #92, #93, #94 and #95.

  • Now building and testing on Windows and MacOS as well as Linux.

  • Python version 3.12 added to package classifiers

  • Python version 3.12 tested during CI

0.4.1, 2 October 2023¶

  • This is a minor release intended to preview new pandas-related features for version 0.5.0.

  • Added another pandas Series accessor, is_imbalanced().

  • Added two pandas DataFrame accessors, feature_importances() and correlation_detector(). These are experimental features.

0.4.0, 28 September 2023¶

  • redflag can now be installed by the conda package and environment manager. To do so, use conda install -c conda-forge redflag.

  • All of the sklearn components can now be instantiated with warn=False in order to trigger a ValueException instead of a warning. This allows you to build pipelines that will break if a detector is triggered.

  • Added redflag.target.is_ordered() to check if a single-label categorical target is ordered in some way. The test uses a Markov chain analysis, applying chi-squared test to the transition matrix. In general, the Boolean result should only be used on targets with several classes, perhaps at least 10. Below that, it seems to give a lot of false positives.

  • You can now pass groups to redflag.distributions.is_multimodal(). If present, the modality will be checked for each group, returning a Boolean array of values (one for each group). This allows you to check a feature partitioned by target class, for example.

  • Added redflag.sklearn.MultimodalityDetector to provide a way to check for multimodal features. If y is passed and is categorical, it will be used to partition the data and modality will be checked for each class.

  • Added redflag.sklearn.InsufficientDataDetector which checks that there are at least M2 records (rows in X), where M is the number of features (i.e. columns) in X.

  • Removed RegressionMultimodalDetector. Use MultimodalDetector instead.

0.3.0, 21 September 2023¶

  • Added some accessors to give access to redflag functions directly from pandas.Series objects, via an ‘accessor’. For example, for a Series s, one can call minority_classes = s.redflag.minority_classes() instead of redflag.minority_classes(s). Other functions include imbalance_degree(), dummy_scores() (see below). Probably not very useful yet, but future releases will add some reporting functions that wrap multiple Redflag functions. This is an experimental feature and subject to change.

  • Added a Series accessor report() to perform a range of tests and make a small text report suitable for printing. Access for a Series s like s.redflag.report(). This is an experimental feature and subject to change.

  • Added new documentation page for the Pandas accessor.

  • Added redflag.target.dummy_classification_scores(), redflag.target.dummy_regression_scores(), which train a dummy (i.e. naive) model and compute various relevant scores (MSE and R2 for regression, F1 and ROC-AUC for classification tasks). Additionally, both most_frequent and stratified strategies are tested for classification tasks; only the mean strategy is employed for regression tasks. The helper function redflag.target.dummy_scores() tries to guess what kind of task suits the data and calls the appropriate function.

  • Moved redflag.target.update_p() to redflag.utils.

  • Added is_imbalanced() to return a Boolean depending on a threshold of imbalance degree. Default threshold is 0.5 but the best value is up for debate.

  • Removed utils.has_low_distance_stdev.

0.2.0, 4 September 2023¶

  • Moved to something more closely resembling semantic versioning, which is the main reason this is version 0.2.0.

  • Builds and tests on Python 3.11 have been successful, so now supporting this version.

  • Added custom ‘alarm’ Detector, which can be instantiated with a function and a warning to emit when the function returns True for a 1D array. You can easily write your own detectors with this class.

  • Added make_detector_pipeline() which can take sequences of functions and warnings (or a mapping of functions to warnings) and returns a scikit-learn.pipeline.Pipeline containing a Detector for each function.

  • Added RegressionMultimodalDetector to allow detection of non-unimodal distributions in features, when considered across the entire dataset. (Coming soon, a similar detector for classification tasks that will partition the data by class.)

  • Redefined is_standardized (deprecated) as is_standard_normal, which implements the Kolmogorov–Smirnov test. It seems more reliable than assuming the data will have a mean of almost exactly 0 and standard deviation of exactly 1, when all we really care about is that the feature is roughly normal.

  • Changed the wording slightly in the existing detector warning messages.

  • No longer warning if y is None in, eg, ImportanceDetector, since you most likely know this.

  • Some changes to ImportanceDetector. It now uses KNN estimators instead of SVMs as the third measure of importance; the SVMs were too unstable, causing numerical issues. It also now requires that the number of important features is less than the total number of features to be triggered. So if you have 2 features and both are important, it does not trigger.

  • Improved is_continuous() which was erroneously classifying integer arrays with many consecutive values as non-continuous.

  • Note that wasserstein no longer checks that the data are standardized; this check will probably return in the future, however.

  • Added a Tutorial.ipynb notebook to the docs.

  • Added a Copy button to code blocks in the docs.

0.1.10, 21 November 2022¶

  • Added redflag.importance.least_important_features() and redflag.importance.most_important_features(). These functions are complementary (in other words, if the same threshold is used in each, then between them they return all of the features). The default threshold for importance is half the expected value. E.g. if there are 5 features, then the default threshold is half of 0.2, or 0.1. Part of Issue 2.

  • Added redflag.sklearn.ImportanceDetector class, which warns if 1 or 2 features have anomalously high importance, or if some features have anomalously low importance. Part of Issue 2.

  • Added redflag.sklearn.ImbalanceComparator class, which learns the imbalance present in the training data, then compares what is observed in subsequent data (evaluation, test, or production data). If there’s a difference, it throws a warning. Note: it does not warn if there is imbalance present in the training data; use ImbalanceDetector for that.

  • Added redflag.sklearn.RfPipeline class, which is needed to include the ImbalanceComparator in a pipeline (because the common-or-garden sklearn.pipeline.Pipeline class does not pass y into a transformer’s transform() method). Also added the redflag.sklearn.make_rf_pipeline() function to help make pipelines with this special class. These components are straight-up forks of the code in scikit-learn (3-clause BSD licensed).

  • Added example to docs/notebooks/Using_redflag_with_sklearn.ipynb to show how to use these new objects.

  • Improved redflag.is_continuous(), which was buggy; see Issue 3. It still fails on some cases. I’m not sure a definitive test for continuousness (or, conversely, discreteness) is possible; it’s just a heuristic.

0.1.9, 25 August 2022¶

  • Added some experimental sklearn transformers that implement various redflag tests. These do not transform the data in any way, they just inspect the data and emit warnings if tests fail. The main ones are: redflag.sklearn.ClipDetector, redflag.sklearn.OutlierDetector, redflag.sklearn.CorrelationDetector, redflag.sklearn.ImbalanceDetector, and redflag.sklearn.DistributionComparator.

  • Added tests for the sklearn transformers. These are in redflag/tests/test_redflag.py file, whereas all other tests are doctests. You can run all the tests at once with pytest; coverage is currently 94%.

  • Added docs/notebooks/Using_redflag_with_sklearn.ipynb to show how to use these new objects in an sklearn pipeline.

  • Since there’s quite a bit of sklearn code in the redflag package, it is now a hard dependency. I removed the other dependencies because they are all dependencies of sklearn.

  • Added redflag.has_outliers() to make it easier to check for excessive outliers in a dataset. This function only uses Mahalanobis distance and always works in a multivariate sense.

  • Reorganized the redflag.features module into new modules: redflag.distributions, redflag.outliers, and redflag.independence. All of the functions are still imported into the redflag namespace, so this doesn’t affect existing code.

  • Added examples to docs/notebooks/Basic_usage.ipynb.

  • Removed the class_imbalance() function, which was confusing. Use imbalance_ratio() instead.

0.1.8, 8 July 2022¶

  • Added Wasserstein distance comparisons for univariate and multivariate distributions. This works for either a groups array, or for multiple dataset splits if that’s more convenient.

  • Improved get_outliers(), removing OneClassSVM method and adding EllipticEnvelope and Mahalanobis distance.

  • Added Mahalanobis distance outlier detection function to serve get_outliers() or be used on its own. Reproduces the results zscore_outliers() used to give for univariate data, so removed that.

  • Added kde_peaks() function to find peaks in a kernel density estimate. This also needed some other functions, including fit_kde(), get_kde(), find_large_peaks(), and the bandwidth estimators, bw_silverman() and bw_scott().

  • Added classes argument to the class imbalance function, in case there are classes with no data, or to override the classes in the data.

  • Fixed a bug in the feature_importances() function.

  • Fixed a bug in the is_continuous() function.

  • Improved the Using_redflag.ipynb notebook.

  • Added has_nans(), has_monotonic(), and has_flat() functions to detect interpolation issues.

  • Moved some more helper functions into utils, eg iter_groups(), ecdf(), flatten(), stdev_to_proportion() and proportion_to_stdev().

  • Wrote a lot more tests, coverage is now at 95%.

0.1.3 to 0.1.7, 9–11 February 2022¶

  • Added utils.has_low_distance_stdev.

  • Added utils.has_few_samples.

  • Added utils.is_standardized() function to test if a feature or regression target appears to be a Z-score.

  • Changed name of clips() function to clipped() to be more predictable (it goes with is_clipped()).

  • Documentation.

  • CI workflow seems to be stable.

  • Mostly just a lot of flailing.

0.1.2, 1 February 2022¶

  • Early release.

  • Added auto-versioning.

0.1.1, 31 January 2022¶

  • Early release.

0.1.0, 30 January 2022¶

  • Early release.