Robust methods for high-dimensional data, in particular linear model selection techniques based on least angle regression and sparse regression.
Changes in robustHD version 0.5.1
+ Explicitly calling C++ function std::abs() rather than abs() to avoid
clang warning.
+ Correctly importing functions head() and tail() from package 'utils' and
function devAskNewPage from package 'grDevices'.
Changes in robustHD version 0.5.0
+ Added functionality for (robust) groupwise least angle regression.
+ Added TopGear car data.
+ Diagnostic plots now allow to pass arguments to covMcd().
+ Removed PCA step from data cleaning RLARS to consolidate code.
+ Updated package dependencies.
Changes in robustHD version 0.4.0
+ sparseLTS() no longer uses subsampling algorithm in the special case of
alpha = 1.
+ sparseLTS() now has argument 'normalize' to specify whether the predictor
variables should be normalized.
+ sparseLTS() now computes objective function with coefficients for
normalized data (if applicable).
+ Most required packages are now imports rather than depends.
Changes in robustHD version 0.3.2
+ Bugfixes in sparseLTS() preventing errors for high-dimensional data.
Changes in robustHD version 0.3.1
+ rlars now uses perryFit() instead of perryTuning() for prediction error
estimation.
+ Bugfix in rlars() allowing the number of variables to be sequenced to be
larger than half the number of observations.
+ Bugfix in sparseLTS() in case of only one predictor variable.
+ Added tests for C++ implementation of the lasso.
Changes in robustHD version 0.3.0
+ Redesign of the class structure.
+ Redesign of how C++ back end is called.
+ Functionality of sparseLTSGrid() now included in sparseLTS();
sparseLTSGrid() is now a deprecated wrapper function.
+ Restructured internal code for computing initial subsets for sparse LTS.
+ rlars() now supports data cleaning RLARS, with an extra PCA step for
high-dimensional data.
+ New argument 's' in rlars() to select the steps along the sequence for
which to compute submodels
+ fortify() and diagnosticPlot() methods for class "seqModel".
+ Bugfix in predict() method for "sparseLTS" if object was computed without
intercept.
Changes in robustHD version 0.2.2
+ Bugfix in sparseLTS() for more stability of the results.
+ Bugfix in winsorize(): weights are now correctly returned as vector for
a matrix with only one column.
+ Bugfix in diagnosticPlot(): previous setting of devAskNewPage() is now
retained on exit.
Changes in robustHD version 0.2.1
+ Bugfix in rlars(): formula method now only adds function call and model
terms if the default method returns an "rlars" object, not if only the
sequence is returned.
+ Bugfix in rlars(): argument cl is now preferred over argument ncores for
parallel computing, as stated in the help file.
+ Plots are no longer using the opts() function from package ggplot2, which
is deprecated since ggplot2 version 0.9.2.
Changes in robustHD version 0.2.0
+ Graphics are now based on package ggplot2 instead of lattice.
+ Prediction error estimation is now based on package perry instead of
cvTools.
+ Parallel computing for sparseLTS() now available via OpenMP.
+ rlars() is now using C++ code for variable sequencing, including
parallelization of certain tasks via OpenMP. Further parallel
computing is implemented on the R level via package parallel.
+ sparseLTSGrid() and rlars() now allow model selection based on the
prediction error.
+ coef(), fitted(), residuals() and wt() methods now have argument
'drop' to control whether to reduce the dimension if possible.
+ Renamed components 'weight' and 'raw.weights' of sparse LTS models to
'wt' and 'raw.wt', and renamed the accessor function accordingly to wt().
+ Print methods for "sparseLTS" and "sparseLTSGrid" now only show non-zero
coefficients by default; also added argument to print method for "rlars".
+ sparseLTS() and sparseLTSGrid() now store the raw fitted values.
+ Bugfixes in C++ code for sparseLTS() and fastLasso() to prevent memory
related errors.