v0.0.23

Improve

• Add UMAP
• Add predict function in Sammon, SNE and tSNE
• Improve LLE process
• Add Zero-inflated poisson model
• Add Bilinear, Cubic convolution, Lanczos and Sinc interpolation
• Add Support vector clustering
• Add Diffusion map
• Add trigonometric and hyperbolic function layers

Breaking changes

• Change FastMap, Isomap, LaplacianEigenmaps, LLE, LSA, MDS and RandomProjection functions to classes
• Change Latent Dirichlet Allocation model to accept categorical data
• Change arguments of functions of Isolation forest
• Change argument type of predict function of SVM, SVR and OCSVM
• Split PAM and CLARA into separate files

Minor changes

• predict function of PercentileAnormaly returns boolean values
• Remove unused variables in A2C model

v0.0.22

Improve

• Add Fuzzy k-nearest neighbor
• Add radius neighbor models
• Add Additive coupling layer and NICE model
• Add model evaluate functions

Breaking changes

• Change retun value of predicted function of NAROW and Confidence weighted
• Change return value of Gaussian process
• Change argument and return type of Polynomial interpolation
• Change I/F of VAE

Bug fix

• weighted method of Lagrange interpolation returns invalid values
• GAN failed to learn when a dimension of input data is not 2
• Maximum likelihood returns invalid values
• rest function of Mountaincar environment returns invalid values

Document

• Add link to source
• Improve many JSDoc

v0.0.21

Improve

• Add Standardizes Major Axis regression and Deming regression
• Add Huber and Tukey regression
• Export Graph and Layer class for Neuralnetwork
• Add Repeated median regression
• Add Self-training model
• Add Delaunay interpolation
• Add Complement, Negation, Universal-set and Selective Naive Bayes
• Add probability method in NaiveBayes class

Breaking change

• Fix class name PassingBablock to PassingBablok
• Split MLP model to classifier and regressor, and specializes in that function
• AODE, CRF, HMMClassifier returns null if not be estimated for any label
• Improve NaiveBayes class to accept distribution at constructor
• Split RandomForest model to classifier and regressor
• Unify interface of classification model to accept target data as Array of any type

Bug fix

• Weight of LeastAbsolute and LpNormLinearRegression are using averages

Minor change

• Change KNN class structure
• Improve structure of AODE class

v0.0.20

Improve

• Add Metropolis-Hastings algorithm
• Add Bayesian Network
• Add Passing-Bablok method

Breaking changes

• predict function of FuzzyCMeans, PossibilisticCMeans and SoftKMeans returns 2D array represents each categories

Bug fix

• A2C throw error sometime
• Use of Array.sort function in RobustScaler is incorrect

Minor changes

• Change structure of NeuralNetwork to separate graph structure and layer calculation
• copy function of NeuralNetwork do not use fromObject function of NeuralNetwork
• Change boolean operation of Thompson

v0.0.19

Improve

• Add llh to GPLVM
• Some argument of kernel/kernelArgs becomes unnecessary
• Add k-modes and k-prototypes

Breaking changes

• Change to default export of SplineInterpolation and SmoothingSpline
• Change to default export of HoltWinters
• Fix name of JeoJohnson to YeoJohnson
• Remove env parameter from some unnecessary functions

Bug fix

• predict function of HoltWinters
• Parameters of HMM become NaN sometime

Minor changes

• Change default parameters of GPLVM
• Change default parameters of PPR
• Improve initialize of ARMA
• Fix JSDoc
• Create command of creating .d.ts file

v0.0.18

Improve

• Export Complex number class
• Add CAST clustering
• Add LMNN classification
• Add KernelPCA and DualPCA model
• Add scaling methods (MaxAbsolute, MinMax, Robust, Standardization)
• Add ddof parameter to variance and std function of Matrix class
• Add Probabilistic PCA and GPLVM

Breaking changes

• Change name hierarchy to agglomerative
• HMM only accept categorical data (same as CRF)
• Remove kernel computation from PCA

Bug fix

• predict function of HMMClassifier always returns -1
• The arguments of callback function for reduce of Matrix class is invalid when axis is -1
• The calculation of percentile is invalid in some models
• predict function of KNN, KNNRegression, KNNAnomaly and KNNDensityEstimation only accept one data

Minor changes

• Change learning rate of CRF
• Split some algorithms (Lasso and Lagrange interpolation)

Documents

• Change notation of unnecessary parameters

v0.0.17

Improve

• Add CRF
• Change initial values of hessenbergArnoldi function of matrix class
• Add toScaler function of matrix class
• row and col of matrix class accept Array of boolean
• Add LSIF

Breaking change

• Change arguments of callback function of reduce

Bug fix

• Failed when apply Array class to predict function in KLIEP, uLSIF, RuLSIF and LSDD
• fit of KLIEP and RuLSIF is not correct
• Polynomial kernel of Kernel PCA is incorrect
• Second and third arguments of callback function of every and some is undefined

v0.0.16

Improve

• Add Gasser–Müller kernel estimator
• Add predict function in hierarch clustering methods
• Add weighted least squares
• Add block function of Matrix class

Breaking change

• Make null as unlabeled data of semi-supervised learning models
• Change slice and remove sliceRow and sliceCol function of Matrix class
• Change property name of Affinity propagation
• predict function returns 1 or -1 of some binary classification models

v0.0.15

Improve

• Add G-means
• Change to manage optimizer in neuralnetwork
• Add levinson and householder method in AR model
• Add VAR model
• Remove dependency of HoltWinters in ExponentialAverage
• Separate RNN units from the layer
• at and set function of Matrix and Tensor accept Array at the first argument

Bug fix

• Invalid addition in AR model
• Some models not working because of the arguments of callback function in Matrix changed (Label propagation, Label spreading, Laplacian eigenmaps, Cond layer, Huber layer)
• predict of Probit returns not Array
• Polynomial histogram throws error

v0.0.14

Improve

• Add batch normalization and reverse layer of neuralnetwork
• Add Incremental PCA
• Accept Matrix instance in diag function
• Export Complex number class
• at function of Tensor throws error when the index is out bounds
• Unify the parameters of callback of forEach and map
• Add kron and rank function in Matrix
• Add ActivationLayer and use in some layers
• Layer naming done in Layer class
• Add fromObject and toObject functions in Layer/Matrix class

Breaking change

• Removing the KMeans dependency of Neural Gas
• Change the constructor argument of PCA from a function to a string
• Remove most of dst parameters of functions of Matrix
• Remove getLayerFromName function from NeuralNetwork
• Remove get_params and set_params from Layer classes

Bug fix

• predict function of GBDT and XGBoost throw error
• sqrt function of Complex returns three values
• quantile function of Matrix returns invalid values
• If the kernel parameter of convolute function is not square, the function returns incorrect values