0.12.0 - 2022-09-02

• Moved all the public modules imports from river/__init__.py to river/api.py and removed unnecessary dependencies between modules enabling faster cherry-picked import times (~3x).
• Adding wheels for Python 3.11.

base

• Introduced an mutate method to the base.Base class. This allows setting attributes in a controlled manner, which paves the way for online AutoML. See the recipe for more information.

compat

• Moved the PyTorch wrappers to river-extra.

covariance

• Created a new covariance module to hold everything related to covariance and inversion covariance matrix estimation.
• Moved misc.CovarianceMatrix to covariance.EmpiricalCovariance.
• Added covariance.EmpiricalPrecision to estimate the inverse covariance matrix.

compose

• Moved utils.pure_inference_mode to compose.pure_inference_mode and utils.warm_up_mode to compose.warm_up_mode.
• Pipeline parts can now be accessed by integer positions as well as by name.

datasets

• Imports synth, enabling `from river import datasets; datasets.synth.

drift

• Refactor the concept drift detectors to match the remaining of River's API. Warnings are only issued by detectors that support this feature.
• Drifts can be assessed via the property drift_detected. Warning signals can be acessed by the property warning_detected. The update now returns self.
• Ensure all detectors automatically reset their inner states after a concept drift detection.
• Streamline DDM, EDDM, HDDM_A, and HDDM_W. Make the configurable parameters names match their respective papers.
• Fix bugs in EDDM and HDDM_W.
• Enable two-sided tests in PageHinkley.
• Improve documentation and update tests.

feature_extraction

• Added a tokenizer_pattern parameter to feature_extraction.BagOfWords and feature_extraction.TFIDF to override the default pattern used for tokenizing text.
• Added a stop_words parameter to feature_extraction.BagOfWords and feature_extraction.TFIDF for removing stop words once the text has been tokenized.

linear_model

• After long ado, we've finally implemented linear_model.BayesianLinearRegression.

metrics

• Removed dependency to optim.
• Removed metrics.Rolling, due to the addition of utils.Rolling.
• Removed metrics.TimeRolling, due to the addition of utils.Rolling.

proba

• Removed proba.Rolling, due to the addition of utils.Rolling.
• Removed proba.TimeRolling, due to the addition of utils.Rolling.

rule

• The default splitter was changed to tree.splitter.TEBST for memory and running time efficiency.

stats

• Removed stats.RollingMean, due to the addition of utils.Rolling.
• Removed stats.RollingVar, due to the addition of utils.Rolling.
• Removed stats.RollingCov, due to the addition of utils.Rolling.
• Removed stats.RollingPearsonCorr, due to the addition of utils.Rolling.

stream

• stream.iter_array now handles text data.
• Added stream.TwitterLiveStream, to listen to a filtered live stream of Tweets.

time_series

• Added time_series.HorizonAggMetric.
• Fixed a bug in time_series.SNARIMAX where the number of seasonal components was not correct when sp or sq were specified.
• Fixed the differencing logic in time_series.SNARIMAX when d or sd were specified.

tree

• Rename split_confidence and tie_threshold to delta and tau, respectively. This way, the parameters are not misleading and match what the research papers have used for decades.
• Refactor HoeffdingAdaptiveTree{Classifier,Regressor} to allow the usage of any drift detector. Expose the significance level of the test used to switch between subtrees as a user-defined parameter.
• Correct test used to switch between foreground and background subtrees in HoeffdingAdaptiveTreeRegressor. Due to the continuous and unbounded nature of the monitored errors, a z-test is now performed to decide which subtree to keep.
• The default leaf_prediction value was changed to "adaptive", as this often results in the smallest errors in practice.
• The default splitter was changed to tree.splitter.TEBST for memory and running time efficiency.

utils

• Removed dependencies to anomaly and compose.
• Added utils.Rolling and utils.TimeRolling, which are generic wrappers for computing over a window (of time).
• Use binary search to speed-up element removal in utils.SortedWindow.