fastmath.ml.regression — generateme/fastmath 3.0.0-alpha4

OLS, WLS and GLM regression models with analysis.

->family clj

(->family family-map)

(->family variance residual-deviance)

(->family default-link
          variance
          initialize
          residual-deviance
          aic
          quantile-residuals-fun
          dispersion)

Create Family record.

Arguments:

• default-link - canonical link function, default: :identity
• variance - variance function in terms of mean
• initialize - initialization of glm, default: the same as in :gaussian
• residual-deviance - calculates residual deviance
• aic - calculates AIC, default (constantly ##NaN)
• quantile-residuals-fun - calculates quantile residuals, default as in :gaussian
• disperation - value or :estimate (default), :pearson or :mean-deviance

Initialization will be called with ys and weights and should return:

• ys, possibly changed if any adjustment is necessary
• init-mu, starting point
• weights, possibly changes or orignal
• (optional) any other data used to calculate AIC

AIC function should accept: ys, fitted, weights, deviance, observation, rank (fitted parameters) and additional data created by initialization

Minimum version should define variance and residual-deviance.

->link clj

(->link link-map)

(->link g mean mean-derivative)

Creates link record.

Args:

• g - link function
• mean - mean, inverse link function
• mean-derivative - derivative of mean

->string clj multimethod

add-penalty clj

(add-penalty xss penalty penalty-param)

(add-penalty xss penalty penalty-param intercept?)

Adds rows with penalty data to a given seq of seqs (a row matrix).

Penalties and parameters map:

• :ridge - diagonal matrix with sqrt(:lambda) parameter, used in regularized regression.
• :diffs - differences of given :order multiplied by sqrt(:lambda), used in penalized b-splines.

Two add more than one penalty, create seq of penalty types and parameters, eg: [:diffs :ridge]

analysis clj

(analysis model)

Influence analysis, laverage, standardized and studentized residuals, correlation.

b-spline-transformer clj

(b-spline-transformer xs nseg degree)

(b-spline-transformer xl xr nseg degree)

cir clj

(cir ys)

(cir xs ys)

(cir xs ys order)

(cir xs ys ws order)

Centered Isotonic Regression.

Returns shrinked [xs,ys] pair.

Arguments:

• xs - regressor variable
• ys - response variable
• ws - weights (optional)
• order - :asc or :increasing (default), :desc or :decreasing, :non-decreasing and :non-increasing.

dose clj

(dose glm-model)

(dose glm-model p)

(dose glm-model p coeff-id)

(dose {:keys [link-fun xtxinv coefficients]} p intercept-id coeff-id)

Predict Lethal/Effective dose for given p (default: p=0.5, median).

• intercept-id - id of intercept, default: 0
• coeff-id is the coefficient used for calculating dose, default: 1

families clj

family-with-link clj

(family-with-link family)

(family-with-link family params)

(family-with-link family link params)

Returns family with a link as single map.

glm clj

(glm ys xss)

(glm ys
     xss
     {:keys [max-iters tol epsilon family link weights alpha offset
             dispersion-estimator intercept? init-mu simple? transformer names
             decomposition augmentation augmentation-params]
      :or {max-iters 25
           tol 1.0E-8
           epsilon 1.0E-8
           family :gaussian
           alpha 0.05
           intercept? true
           simple? false
           decomposition :cholesky}
      :as params})

Fit a generalized linear model using IRLS method.

Arguments:

• ys - response vector
• xss - terms of systematic component
• optional parameters

Parameters:

• :tol - tolerance for matrix decomposition (SVD and Cholesky), default: 1.0e-8
• :epsilon - tolerance for IRLS (stopping condition), default: 1.0e-8
• :max-iters - maximum numbers of iterations, default: 25
• :weights - optional weights
• :offset - optional offset
• :alpha - significance level, default: 0.05
• :intercept? - should intercept term be included, default: true
• :init-mu - initial response vector for IRLS
• :simple? - returns simplified result
• :dispersion-estimator - :pearson, :mean-deviance or any number, replaces default one.
• :family - family, default: :gaussian
• :link - link
• :nbinomial-theta - theta for :nbinomial family, default: 1.0.
• :transformer - an optional function which will be used to transform systematic component xs before fitting and prediction
• :names - an optional vector of names to use when printing the model
• :decomposition - which matrix decomposition use to find solution, :cholesky (default), :rrqr (rank revealing) or :qr
• :augmentation and augmentation-params - regularization by data augmentation
  • :ridge - adds ridge regresion penalty (intercept is not penalized), default parameters {:lambda 0.1}
  • :diffs - adds differences penalty, use with b-spline-transformation for smoothing, default parameters {:lambda 1.0 :order 2}

Family is one of the: :gaussian (default), :binomial, :quasi-binomial, :poisson, :quasi-poisson, :gamma, :inverse-gaussian, :nbinomial, custom Family record (see ->family) or a function returning Family (accepting a map as an argument)

Link is one of the: :probit, :identity, :loglog, :sqrt, :inverse, :logit, :power, :nbinomial, :cauchit, :distribution, :cloglog, :inversesq, :log, :clog, custom Link record (see ->link) or a function returning Link (accepting a map as an argument)

Notes:

• SVD decomposition is used instead of more common QR
• intercept term is added implicitely if intercept? is set to true (by default)
• :nbinomial family requires :nbinomial-theta parameter
• Each family has its own default (canonical) link.

Returned record implementes IFn protocol and contains:

• :model - set to :glm
• :intercept? - whether intercept term is included or not
• :xtxinv - (X^T X)^-1
• :intercept - intercept term value
• :beta - vector of model coefficients (without intercept)
• :coefficients - coefficient analysis, a list of maps containing :estimate, :stderr, :t-value, :p-value and :confidence-interval
• :weights - weights, :weights (working) and :initial
• :residuals - a map containing :raw, :working, :pearsons and :deviance residuals
• :fitted - fitted values for xss
• :df - degrees of freedom: :residual, :null and :intercept
• :observations - number of observations
• :deviance - deviances: :residual and :null
• :dispersion - default or calculated, used in a model
• :dispersions - :pearson and :mean-deviance
• :family - family used
• :link - link used
• :link-fun - link function, g
• :mean-fun - mean function, g^-1
• :q - (1-alpha/2) quantile of T or Normal distribution for residual degrees of freedom
• :chi2 and :p-value - Chi-squared statistic and respective p-value
• :ll - a map containing log-likelihood and AIC/BIC (GLM and based on deviance, dev+2ED)
• :analysis - laverage, residual and influence analysis - a delay
• :iters and :converged? - number of iterations and convergence indicator
• :decomposition - decomposition used

Analysis, delay containing a map:

• :residuals - :standardized and :studentized residuals (pearsons and deviance)
• :laverage - :hat, :sigmas and laveraged :coefficients (leave-one-out)
• :influence - :cooks-distance, :dffits, :dfbetas and :covratio
• :influential - list of influential observations (ids) for influence measures
• :correlation - correlation matrix of estimated parameters

glm-nbinomial clj

(glm-nbinomial ys xss)

(glm-nbinomial ys
               xss
               {:keys [nbinomial-theta max-iters epsilon]
                :or {max-iters 25 epsilon 1.0E-8}
                :as params})

Fits theta for negative binomial glm in iterative process.

Returns fitted model with :nbinomial-theta key.

Arguments and parameters are the same as for glm.

Additional parameters:

• :nbinomial-theta - initial theta used as a starting point for optimization.

links clj

lm clj

(lm ys xss)

(lm ys
    xss
    {:keys [tol weights alpha intercept? offset transformer names decomposition
            augmentation augmentation-params]
     :or {tol 1.0E-8 alpha 0.05 intercept? true decomposition :cholesky}})

Fit a linear model using ordinary (OLS) or weighted (WLS) least squares.

Arguments:

• ys - response vector
• xss - terms of systematic component
• optional parameters

Parameters:

• :tol - tolerance for matrix decomposition (SVD and Cholesky/QR decomposition), default: 1.0e-8
• :weights - optional weights for WLS
• :offset - optional offset
• :alpha - significance level, default: 0.05
• :intercept? - should intercept term be included, default: true
• :transformer - an optional function which will be used to transform systematic component xs before fitting and prediction
• :names - sequence or string, used as name for coefficient when pretty-printing model, default 'X'
• :decomposition - which matrix decomposition use to find solution, :cholesky (default), :rrqr (rank revealing) or :qr
• :augmentation and augmentation-params - regularization by data augmentation
  • :ridge - adds ridge regresion penalty (intercept is not penalized), default parameters {:lambda 0.1}
  • :diffs - adds differences penalty, use with b-spline-transformation for smoothing, default parameters {:lambda 1.0 :order 2}

Notes:

• SVD decomposition is used instead of more common QR
• intercept term is added implicitely if intercept? is set to true (by default)
• Two variants of AIC/BIC are calculated, one based on log-likelihood, second on RSS/n

Returned record implementes IFn protocol and contains:

• :model - :ols or :wls
• :intercept? - whether intercept term is included or not
• :xtxinv - (X^T X)^-1
• :intercept - intercept term value
• :beta - vector of model coefficients (without intercept)
• :coefficients - coefficient analysis, a list of maps containing :estimate, :stderr, :t-value, :p-value and :confidence-interval
• :weights - initial weights
• :residuals - a map containing :raw and :weighted residuals, also :loocv
• :fitted - fitted values for xss
• :df - degrees of freedom: :residual, :model and :intercept
• :observations - number of observations
• :r-squared and :adjusted-r-squared
• :sigma and :sigma2 - deviance and variance
• :msreg - regression mean squared
• :rss, :regss, :tss - residual, regression and total sum of squares
• :qt - (1-alpha/2) quantile of T distribution for residual degrees of freedom
• :f-statistic and :p-value - F statistic and respective p-value
• :ll - a map containing log-likelihood and AIC/BIC in two variants: based on log-likelihood and RSS
• :analysis - laverage, residual and influence analysis - a delay
• :decomposition - decomposition used
• :augmentation - augmentation used
• :cv - cross validation statistic
• :effective-dimension - effective dimension of the model (sum of hat matrix diagonal)

Analysis, delay containing a map:

• :residuals - :standardized and :studentized weighted residuals
• :laverage - :hat, :sigmas and laveraged :coefficients (leave-one-out)
• :influence - :cooks-distance, :dffits, :dfbetas and :covratio
• :influential - list of influential observations (ids) for influence measures
• :correlation - correlation matrix of estimated parameters
• :normality - residuals normality tests: :skewness, :kurtosis, :durbin-watson (for raw and weighted), :jarque-berra and :omnibus (normality)

pava clj

(pava ys)

(pava ys order)

(pava ys ws order)

Isotonic regression, pool-adjacent-violators algorithm with up-and-down-blocks variant.

Isotonic regression minimizes the (weighted) L2 loss function with a constraint that result should be monotonic (ascending or descending).

Arguments:

• ys - response variable data
• ws - weights (optional)
• order - :asc or :increasing (default), :desc or :decreasing, :non-decreasing and :non-increasing

Returns monotonic predicted values.

polynomial-transformer clj

(polynomial-transformer xs degree)

Creates polynomial transformer for xs

predict clj

(predict model xs)

(predict model xs stderr?)

Predict from the given model and data point.

If stderr? is true, standard error and confidence interval is added. If model is fitted with offset, first element of data point should contain provided offset.

Expected data point:

• [x1,x2,...,xn] - when model was trained without offset
• [offset,x1,x2,...,xn] - when offset was used for training
• [] or nil - when model was trained with intercept only
• [offset] - when model was trained with intercept and offset

quantile-residuals clj

(quantile-residuals {:keys [quantile-residuals-fun residuals dispersion]
                     :as model})

Quantile residuals for a model, possibly randomized.

trigonometric-transformer clj

(trigonometric-transformer period degree)

Creates trigonometric transformer for xs