0.3.1

• Fix CART feature importances purity increase overflow

0.3.0

• Added K Best feature selector
• Added Flysystem 2.0 Persister
• Stateful and Elastic Transformers are now Persistable
• Added Gzip serializer for Persistable objects
• Added Sentence tokenizer
• Library now throws Rubix\ML namespaced exceptions
• Added Scoring interface for estimators that score samples
• Deprecated the Ranking interface
• Add generic Trainable interface
• Decision Trees are now iterable
• Added K-Skip-N-Gram tokenizer and deprecated Skip Gram
• Single sample inference methods are now marked internal
• Deprecated Variance Threshold Filter

Note: This update includes changes to class definitions which may break previously persisted models if upgraded. See the caveats section of the persistence documentation for more info.

0.2.4

• Categorized and annotated internal API
• Fix context of preprocess() and combinations() methods
• Added version constants

0.2.3

• Now compatible with PHP 8 GD Image types
• Dataset cast sample to array upon validation

0.2.2

• Optimized CART quantile-based node splitting
• Fixed CART and Extra Tree min purity increase post pruning
• Fix ITree infinite loop splitting same samples

0.2.1

• Optimized Stop Word Filter
• Allow list of empty regex patterns in Regex Filter
• Handle missing class definitions in Native and Igbinary
• Fixed infinite loop in Ball Tree & KD Tree grow method

0.2.0

• Add Recursive Feature Eliminator feature selector
• Can now disable holdout validation in MLP learners
• TF-IDF Transformer additive (Laplace) smoothing now variable
• Added instability detection to gradient-based learners
• Gradient Boost validation set holdout can now be 0
• Specifications now extend base class
• Rename Dataset validate argument to verify
• Ball Tree Cluster nodes are now called Cliques
• ITree cells are now called Depth nodes
• Added Dataset join() method and deprecated augment()
• Added score() method to Ranking API and deprecated rank()
• Renamed Radius Neighbors anomalyClass to outlierClass
• HTML Stripper can now allow user-specified tags
• Sparse Random Projector now has variable sparsity
• Deprecated Dense Random Projector transformer

Note: This update includes changes to class definitions which may break previously persisted models if upgraded. See the caveats section of the persistence documentation for more info.

0.1.6

• Fix KNN Imputer spatial tree dependency injection

0.1.5

• Compensate for zero vectors in Cosine kernel
• Fixed KMC2 random threshold calculation
• Fix Naive Bayes divide by zero when smoothing is 0

0.1.4

• Optimized Cosine distance for sparse vectors