covariance 0.1.0.3 → 0.1.0.4

raw patch · 5 files changed

+84/−10 lines, 5 filesdep +glasso

Dependencies added: glasso

Files

• CHANGELOG.md +5/−0
• covariance.cabal +6/−4
• src/Statistics/Covariance.hs +9/−5
• src/Statistics/Covariance/GraphicalLasso.hs +62/−0
• test/Test.hs +2/−1

@@ -5,6 +5,11 @@ ## Unreleased changes  +## 0.1.0.4++-   Graphical lasso (for now only re-export from `glasso` library).++ ## 0.1.0.3  -   Fix tests and docs.

@@ -1,6 +1,6 @@ cabal-version:      2.4 name:               covariance-version:            0.1.0.3+version:            0.1.0.4 synopsis:     Well-conditioned estimation of large-dimensional covariance matrices @@ -26,13 +26,15 @@     exposed-modules:       Statistics.Covariance     other-modules:-      Statistics.Covariance.Types+      Statistics.Covariance.GraphicalLasso+      Statistics.Covariance.Internal.Tools       Statistics.Covariance.LedoitWolf-      Statistics.Covariance.RaoBlackwellLedoitWolf       Statistics.Covariance.OracleApproximatingShrinkage-      Statistics.Covariance.Internal.Tools+      Statistics.Covariance.RaoBlackwellLedoitWolf+      Statistics.Covariance.Types     ghc-options: -Wall -Wunused-packages     build-depends:    base ^>=4.14.3.0+                    , glasso                     , hmatrix                     , statistics                     , vector

@@ -21,6 +21,9 @@     module Statistics.Covariance.RaoBlackwellLedoitWolf,     module Statistics.Covariance.OracleApproximatingShrinkage, +    -- * Gaussian graphical model based estimators+    module Statistics.Covariance.GraphicalLasso,+     -- * Misc     DoCenter (..), @@ -33,6 +36,7 @@ import qualified Data.Vector.Storable as VS import qualified Numeric.LinearAlgebra as L import qualified Numeric.LinearAlgebra.Devel as L+import Statistics.Covariance.GraphicalLasso import Statistics.Covariance.LedoitWolf import Statistics.Covariance.OracleApproximatingShrinkage import Statistics.Covariance.RaoBlackwellLedoitWolf@@ -72,8 +76,8 @@ -- 1.0. The estimated covariance matrix of a scaled data matrix is a correlation -- matrix, which is easier to estimate. scale ::-  -- | Data matrix of dimension NxP, where N is the number of observations, and-  -- P is the number of parameters.+  -- | Sample data matrix of dimension \(n \times p\), where \(n\) is the number+  -- of samples (rows), and \(p\) is the number of parameters (columns).   L.Matrix Double ->   -- | (Means, Standard deviations, Centered and scaled matrix)   (L.Vector Double, L.Vector Double, L.Matrix Double)@@ -86,10 +90,10 @@ -- | Convert a correlation matrix with given standard deviations to original -- scale. rescaleWith ::-  -- | Vector of standard deviations (length P)+  -- | Vector of standard deviations of length \(p\).   L.Vector Double ->-  -- | Correlation matrix.+  -- | Correlation matrix of dimension \(p \times p\).   L.Matrix Double ->-  -- | Covariance matrix.+  -- | Covariance matrix of dimension \(p \times p\).   L.Matrix Double rescaleWith ss = L.mapMatrixWithIndex (\(i, j) x -> x * (ss VS.! i) * (ss VS.! j))

@@ -0,0 +1,62 @@+-- |+-- Module      :  Statistics.Covariance.GraphicalLasso+-- Description :  Graphical lasso+-- Copyright   :  (c) 2021 Dominik Schrempf+-- License     :  GPL-3.0-or-later+--+-- Maintainer  :  dominik.schrempf@gmail.com+-- Stability   :  experimental+-- Portability :  portable+--+-- Creation date: Wed Sep 15 09:23:19 2021.+module Statistics.Covariance.GraphicalLasso+  ( graphicalLasso,+  )+where++import Algorithms.GLasso+import Data.Bifunctor+import qualified Numeric.LinearAlgebra as L++-- | Gaussian graphical model based estimator.+--+-- This function estimates both, the covariance and the precision matrices. It+-- is best suited for sparse covariance matrices.+--+-- For now, this is just a wrapper around 'glasso'.+--+-- See Friedman, J., Hastie, T., & Tibshirani, R., Sparse inverse covariance+-- estimation with the graphical lasso, Biostatistics, 9(3), 432–441 (2007).+-- http://dx.doi.org/10.1093/biostatistics/kxm045.+--+-- Return 'Left' if+--+-- - the regularization parameter is out of bounds \([0, \infty)\).+--+-- - only one sample is available.+--+-- - no parameters are available.+--+-- NOTE: This function may call 'error' due to partial library functions.+graphicalLasso ::+  -- | Regularization or lasso parameter; penalty for non-zero covariances. The+  -- higher the lasso parameter, the sparser the estimated inverse covariance+  -- matrix. Must be non-negative.+  Double ->+  -- | Sample data matrix of dimension \(n \times p\), where \(n\) is the number+  -- of samples (rows), and \(p\) is the number of parameters (columns).+  L.Matrix Double ->+  -- | @Either ErrorString (Covariance matrix, Precision matrix)@.+  Either String (L.Herm Double, L.Herm Double)+graphicalLasso l xs+  | l < 0 = Left "graphicalLasso: Regularization parameter is negative."+  | n < 2 = Left "graphicalLasso: Need more than one sample."+  | p < 1 = Left "graphicalLasso: Need at least one parameter."+  | otherwise =+    Right $+      bimap convert convert $ glasso p (L.flatten $ L.unSym sigma) l+  where+    n = L.rows xs+    p = L.cols xs+    (_, sigma) = L.meanCov xs+    convert = L.trustSym . L.reshape p

@@ -46,7 +46,8 @@ estimators =   [ (ledoitWolf DoCenter, "ledoitWolf"),     (raoBlackwellLedoitWolf, "raoBlackwellLedoitWolf"),-    (oracleApproximatingShrinkage, "oracleApproximatingShrinkage")+    (oracleApproximatingShrinkage, "oracleApproximatingShrinkage"),+    (fmap fst . graphicalLasso 1.0, "graphicalLasso")   ]  unitTests :: TestTree