Skip to content

0.11.0 - 2022-05-28


  • There is now a anomaly.base.SupervisedAnomalyDetector base class for supervised anomaly detection.
  • Added api.anomaly.GaussianScorer, which is the first supervised anomaly detector.
  • There is now a anomaly.base.AnomalyFilter base class for anomaly filtering methods. These allow to classify anomaly scores. They can also prevent models from learning on anomalous data, for instance by putting them as an initial step of a pipeline.
  • Added anomaly.ConstantFilter and QuantileFilter, which are the first anomaly filters.
  • Removed anomaly.ConstantThresholder and anomaly.QuantileThresholder, as they overlap with the new anomaly filtering mechanism.


  • Fixed an issue where the _raw_memory_usage property would spin into an infinite loop if a model's property was an itertools.count.



  • A revert method has been added to stats.Gaussian.
  • A revert method has been added to stats.Multinomial.
  • Added dist.TimeRolling to measure probability distributions over windows of time.


  • Add the PeriodicTrigger detector, a baseline capable of producing drift signals in regular or random intervals.
  • The numpy usage was removed in drift.KSWIN in favor of collections.deque. Appending or deleting elements to numpy arrays imply creating another object.
  • Added the seed parameter to drift.KSWIN to control reproducibility.
  • The Kolmogorov-Smirnov test mode was changed to the default ("auto") to suppress warnings (drift.KSWIN).
  • Unnecessary usage of numpy was also removed in other concept drift detectors.


  • Streamline SRP{Classifier,Regressor}, remove unneeded numpy usage, make SRP variants robust against missing features, and fix bugs.
  • Remove unneeded numpy usage AdaptiveRandomForest{Classifier,Regressor}.


  • Added a iter_progressive_val_score function, which does the same as progressive_val_score, except that it yields rather than prints results at each step, which give more control to the user.




  • neighbors.KNNADWINClassifier and neighbors.SAMKNNClassifier have been deprecated.
  • Introduced neighbors.NearestNeighbors for searching nearest neighbors.
  • Vastly refactored and simplified the nearest neighbors logic.


  • Added proba.Rolling to measure a probability distribution over a window.


  • AMRules's debug_one explicitly indicates the prediction strategy used by each rule.
  • Fix bug in debug_one (AMRules) where prediction explanations were incorrectly displayed when ordered_rule_set=True.


  • Added an iter_evaluate function to trace the evaluation at each sample in a dataset.


  • Fix bug in Naive Bayes-based leaf prediction.
  • Remove unneeded numpy usage in HoeffdingAdaptiveTree{Classifier,Regressor}.


  • A revert method has been added to stats.Var.