0.1.3

• Optimized Cosine distance kernel
• Optimized (NaN) Safe Euclidean distance kernel
• Fixed markedness calculation in Multiclass Breakdown
• Prevent infinite loop during spatial tree path finding

0.1.2

• Fixed Grid Search best hyper-parameters method
• Fixed K Means average loss calculation
• Fixed bootstrap estimators tiny bootstrap sets

0.1.1

• Fixed Image Resizer placeholder image
• Fixed Filesystem no write permissions on instantiation
• Nicer Stringable object string representations
• Do not terminate empty Spatial tree leaf nodes
• Additional Filesystem persister checks
• Nicer Dataset object validation error messages

0.1.0

• CV Report Generators now return Report objects
• Dataset describe methods now return Report objects
• Allow hyphens and apostrophes in Word Tokenizer
• Dataset conversion methods now return an Encoding object
• Encodings are now writeable to disk
• Allow classes to be selected for Confusion Matrix
• Fixed divide by zero in Multiclass Breakdown report
• Changed Random Projector minDimensions default max distortion
• Fixed Naive Bayes user-defined class prior probabilities
• Internal CV Learners now check for sufficient hold out data
• Fixed randomize empty dataset object
• Removed setPersister method from Persistent Model
• Added Dataset Has Dimensionality Specification
• Changed name of Tree max depth parameter to max height
• Fixed F Beta division by zero
• Dataset toCSV and toNDJSON accept optional header
• Nicer Verbose Learner logger output
• Screen Logger uses empty channel name by default

0.1.0-rc5

• Improved logging for Verbose Learners
• Added max document frequency to Word Count Vectorizer
• Whitespace Trimmer is now a separate Transformer
• Text Normalizers no longer remove extra whitespace
• Added extra characters pattern to Regex Filter class constants
• Moved Lambda Function transformer to Extras package
• GaussianNB new class labels during partial train
• Decision Tree print ruleset now accepts a header
• Fixed Variance Threshold Filter drop categorical by default
• Removed AdaBoost return learned sample weights

0.1.0-rc4

• Added Multibyte Text Normalizer transformer
• V Measure now has adjustable beta parameter
• Persistent Model is no longer Verbose
• Stop Word Filter now handles unicode characters

0.1.0-rc3

• Embedders now adopt the Transformer API
• Added RanksFeatures interface
• Logistic Regression and Adaline now implement RanksFeatures
• Ridge now implements the RanksFeatures interface
• Added L2 regularization to Dense hidden layers
• Neural Network L2 regularization now optional
• Added MLP numerical instability checks
• Optimized Ball Tree nearest neighbors search
• Pipeline is now more verbose
• Renamed Dataset partition method to partitionByColumn
• Decreased default neural net learner batch size to 128
• Increased default K Means batch size to 128
• Renamed Dataset types method to columnTypes
• Efficient serialization of Word Count Vectorizer
• Decoupled Persistable interface from Learner
• Moved Gower Distance kernel to Extras package
• Moved SiLU activation function to Extras package
• Removed array_first and array_last from global functions
• Abstracted deferred Backend computations into Tasks
• Removed unused BST interface

0.1.0-rc2

• Persistent Model now implements Verbose interface
• Tuned CART continuous feature quantile-based split finding
• N-gram and SkipGram use configurable base word tokenizer
• Moved Alpha Dropout hidden layer to Extras package
• Added Dataset merge and augment methods
• Removed Dataset prepend and append methods
• Lambda Function transformer now takes any callable
• Text Normalizer trim extra whitespace not optional
• Mean Shift minimum seeds now set at 20
• Standardized K Means inertial loss over batch count
• Added set persister method to Persistent Model
• Removed range() from neural network Cost Function interface
• Increased default neural net learner batch size to 200

0.1.0-rc1

• Random Forest now handles imbalanced datasets
• Added early stopping window to AdaBoost
• Gaussian MLE now has automatic and adaptive threshold
• Loda now has automatic and adaptive threshold
• Variance Threshold Filter now selects top k features
• Added params method to Estimator and Embedder interface
• t-SNE now compatible with categorical distance kernels
• Grid Search implements the Wrapper interface
• Grid Search memoizes all results from last search
• Dataset fromIterator method accepts any iterable
• Column Picker throws exception if column not found
• Better hyper-parameter stringification
• Improved Dataset exception messages
• RMSE now default validation Metric for Regressors
• Added balanced accuracy and threat score to Multi-class report
• Pipeline and Persistent Model now implement Ranking
• Changed percentile to quantile in Stats helper
• Renamed Residual Analysis report to Error Analysis
• Changed namespace of specification objects

0.0.19-beta

• Added SiLU self-stabilizing neural network activation function
• Dense hidden layers now have optional bias parameter
• KNN-based imputers accelerated by spatial tree
• Changed the default anomaly class for Radius Neighbors
• Removed additional methods from guessing Strategies
• Numeric String Converter now uses fixed NaN placeholder
• Missing Data Imputer now passes through other data types
• Changed order of Missing Data Imputer params
• Renamed high-level resource type to image type
• Added comb (n choose k) to global functions
• Image Vectorizer now has grayscale option
• Clusterers and Anomaly Detectors return integer predictions
• Ball Tree now compatible with categorical distance kernels
• Parallel Learners using Amp Backend are now persistable
• Changed order of Radius Neighbors hyper-parameters