Changes in version 1.28.0:
o New fct correlratio()
o Many bugs corrected in fct sensiHSIC()
o Many additional options added in fct sensiHSIC(), especially HSIC-ANOVA indices
o New fct testHSIC() for independence testing based on HSIC indices
o New benchmark matyas.fun() added in "testmodels.R"
Changes in version 1.27.1:
o Adding the topic of 'interpretability of machine learning models'
in the description of the package
o emvd() fct and names have been changed to pmvd()
o Several bugs corrected in pme_knn()
o R requirements: Changes from 'class()' to 'inherits()' in several fcts
o Corrections in parameterSets(),
see https://github.com/cran/sensitivity/pull/2
Changes in version 1.27.0:
o New fct pme_knn() for proportional marginal effects computation
o Addition of an example about visualiz. of perturbed law in PLI() fct
Changes in version 1.26.1:
o Simplification in shapleyBlockestimation.R and
parameter values changes in shapleyBlockestimation.Rd
o extract and ggplot functionalities added in shapleysobol_knn() fct
o new examples for nice visualizations (via plotCI() of plotrix package)
in fcts PLIquantile(), PLIsuperquantile(), PLIquantile_multivar()
and PLIsuperquantile_multivar()
Changes in version 1.26.0:
o New function shapleysobol_knn() which allows optimized and parallelized
computations of Shapley effects/Sobol' indices using the
nearest-neighbors method. This fct also replicates the same functionalities
than sobolshap_knn() that will be removed in a future package's version
o \donttest examples have been changed in the .Rd of several functions
in order to shorten their execution (because the tests are judged too long
by CRAN)
o Bug corrected in all PLIxxx() fcts:
- conditions (delta != 0) replaced by conditions (delta != mu),
mu being the input mean)
- pti=phi(lambda[K]) is not recomputed at each bootstrap iteration
(causing numerical instabilities)
o Bug corrected in PLIquantile_multivar() and PLIsuperquantile_multivar()
fcts (conditions "|" instead of "&")
o Correction in docs of PLIquantile_multivar() and
PLIsuperquantile_multivar() to precise that the delta are perturbed means
o In the fct shapleyPermrand(), the test about 'Multiserver queue model'
has been suppressed (unsolved bug in Xset)
Changes in version 1.25.0:
o New function PLIsuperquantile_multivar()
o Bootstrap added in PLIquantile_multivar()
o Bug corrected in weightTSA() (data structures)
o Change in lmg(): new \code{max.iter} argument to increase the maximum
number of iterative optimization step for logistic regression
o Change in emvd():
- new \code{max.iter} argument to increase the maximum number of
iterative optimization step for logistic regression
- new \code{tol} argument to detect spurious variables
- faster implementation
- recursive emvd computation can now be partially done in parallel
o Add an example with Gaussian inverse transforms in PLI() fct
Changes in version 1.24.1:
o Bug for y format in lmg() and emvd(): allow dataframe
Changes in version 1.24.0:
o New function lmg() for LMG sensitivity indices
o New function emvd() for E-MVD sensitivity indices
o Package foreach is necessary (Imports)
Changes in version 1.23.1:
o Bug of tests of sensiHSIC()
o Bug in sobolshap_knn() (matrix option in argument U)
Changes in version 1.23.0:
o New function qosa() computing quantile-oriented sensitivity analysis
via the contrast-function formulation
o Corrections in sensiHSIC() for
- taking into account categorical inputs (and the choice of their
hyperparameters)
- taking into account multiple outputs when some are categorical
o Corrections in sobolshap_knn() for taking into account categorical
inputs
o Correction in bootstats() relative to the use of sobolrep() fct
Changes in version 1.22.2:
o Correction in PLIsuperquantile() for "bias=TRUE" & "type=VAR" case
o New toy function in testmodels(): Friedman function
o New argument 'varoutput' in delsa() in order to be able to
take the output variance as the empirical variance from the sample
o New argument 'randperm' in sobolshap_knn() in order to use or not
random permutations to estimate Shapley effects
Changes in version 1.22.1:
o Correction in sensiHSIC.R with test.method = "No" in order to
not perform the statistical tests
o Correction in nested.cpp
o Change of colors in shapleyBlockEstimation.Rd
Changes in version 1.22.0:
o New function shapleyBlockEstimation()
o New functionalities in sensiHSIC(): PCA, kernel for categorical
inputs and/or outputs, optimization of permutations numbers,
convergence plot for p-values, U-stat estimator for conditional HSIC
o Corrected plot with new fct plotMultOut() (for ubiquitous Sobol'
indices) in sobolMultOut() and associated sobolxxx() fcts
o New bootstrap process in PLIquantile() and PLIsuperquantile() via
the new argument bootsample
Changes in version 1.21.0:
o New function sobolshap_knn() (sensitivity analysis via
ranking / nearest neighbours)
o New functionalities in sensiHISIC(): aggregated HSIC
Changes in version 1.20.0:
o New functions sobolrep() (Sobol' indices estimation based on
replicated orthogonal arrays) and sobolrec() (Recursive estimation
of Sobol' indices), which use the new functions addelman_const(),
discrepancyCriteria_cplus() and maximin_cplus() (discrepancy and
maximin criteria using a C++ implementation)
o bootstrap implemented in sobolrank()
o New function shapleyLinearGaussian() which computes Shapley effects
for linear models with Gaussian inputs
o bug corrected in shapleySubsetMc() with the modified line:
cost=sum(apply(U,1,function(u) round(Ntot/choose(p,sum(u))/(p-1))))
Changes in version 1.19.0:
o New author Sebastien Da veiga
o New function sobolrank()
o Functions src() and pcc() adapted to work with logistic GLM model
(via the new argument logistic = TRUE)
o New argument 'title' in ggplot.shapleyPermRand and
ggplot.shapleyPermEx
o bug solved in shapleyPermEx() in order to compute the CI for the
full first order and independent total Sobol' indices
o New argument 'noise' to consider noisy observations in
shapleySubsetMc()
Changes in version 1.18.1:
o New argument semi (= TRUE or FALSE) in function pcc() in order
to compute semi-partial correlation coefficient (SPCC)
o bug (caused by new R version 4.0.0) solved in:
- sobolmartinez.R, base.R, morris.R, morris_oat.R and simplex.R
(matrix objects now also inherit from class "array")
- sobolMultOut.Rd, testmodels.Rd and delsa.Rd (screen devices
x11() should not be used in examples)
Changes in version 1.18.0:
o New function: weightTSA.R in order to perform Target Sensitivity
Analysis by transforming the output sample using aspecific weight
function
o Improvements in sensiHSIC(): 1) for the parameters estimation via
the new argument crit.option, 2) for a sequential estimation of the
pvalue (independence test based on permutations), 3) new argument
target in order to perform Target Sensitivity Analysis
o New argument 'bias' in PLIsuperquantile(): "TRUE" gives the
previous computation; "FALSE" corrects this computation (the sample
points were not weighted by the likelihood ratio when computing the
perturbed superquantile)
Changes in version 1.17.1:
o bug (caused by new R version 4.0.0) solved in morris_oat.R
(matrix objects now also inherit from class "array")
o exception "multiIndex[inputIndex] == 0 & der" solved in
PoincareChaosSqCoef.R
o example with given data in sensiHSIc()
Changes in version 1.17.0:
o New functions: squaredIntEstim.R and PoincareChaosSqCoef.R
(which allows to compute the squared coefficients in
generalized chaos via Poincare differential operators)
o Modification in PoincareOptimal.Rd concerning the
returned eigenvalues
Changes in version 1.16.3:
o Unresolved matrix/array classes related bug on R4.0.0 version
in morris.R, morris_oat.R, soboljansen.R, sobolmartinez.R
and sobolroalhs.R
Examples that do not work have been put in \donttest{} (these
examples work in practice)
o Resolved bug in plot.soboljansen() and plot.sobolmartinez
when dealing with matrice-type model outputs
o ylabel changed (from S to PLI) when plotting PLI examples
o legends precised (between full and independent Sobol' indices)
when plotting Shapley effects
Changes in version 1.16.2:
o New functionalities in sensiHSIC() function
(statistical tests of independence)
o Graph legends corrected in sensiHSIC(), sensiFdiv(), sb() and
sobolMultOut() functions
o Corrected bug in nodeggplot()
Changes in version 1.16.1:
o New function nodeggplot() (in order to ebautify the plots)
o ggplot functions for the following functions:
pcc, src, sensiFdiv, sensiHSIC, shapleyPermEx, shapleyPermrand,
sobol, sobol2002, sobol2007, sobolEff, soboljansen, sobolmara,
sobolmartinez, sobolMultOut, sobolowen, sobolroalhs, sobolroauc,
sobolsalt, sobolTIIlo, sobolTIIpf, soboltouati
o Functions sobolroalhs() and sobolSalt(): Modifications of
plot functions in order to have "XiXj" instead of "Xij""
o New output of function delsa(): Brute values of derivatives
o Test case corrected in pcc.Rd
Changes in version 1.16.0:
o Added functions for calculating Shapley indices from data:
shapleySubsetMC()
o Added functions for calculating PLI: PLIsuperquantile() and
PLIquantile_multivar()
o In the 3 PLI functions related to quantiles:
- new arguments in order to have the PLI in percentage
- the result also contains the perturbed quantiles values
- bootstrap confidence intervals can be obtained
o Added the toy function heterdisc.fun()
Changes in version 1.15.2:
o Added functions for Morris method for multidimensional outputs:
morrisMultOut()
o Added the toy function linkletter.fun()
o Change the names of the fcts (dtnorm, ..., dtgumbel, ...) by
(dnorm.trunc, ...) (reverse dependency pb with ATmet package).
Export now the fcts, documentation entries in truncateddistrib.Rd
o Corrected bug: Replace \dontrun by \donttest in all Rd-files
o Corrected bug: package triangle is suggested and called in
PoincareOptimal.Rd
o Function in sobolGPmethods.R and sobolpickfreeze.R have been
integrated in sobolGP.R
o Detected bug (not corrected) in tell.sobolGP()
Unused argument in km(), passed by eval(), update(), tell.sobolGP() ??)
=> comment in the example of sobolGP.Rd
Changes in version 1.15.1:
o Functions shapleyPermEx() corrected (confidence intervals)
Changes in version 1.15.0:
o Function "sobolCert.R" has been (temporarily) suppressed
(compilation bug unresolved)
o Function "support.R" has been changed by O. Roustant (new plotting
functionalities are available, as the function "scatterplot")
o Resolved bug about "sobolmartinez" method (now suppressed) in
the sobolMultOut() function
o sobolMultOut() now returns the functional Sobol' indices via
the attributes "Sfct" and "Tfct" of the output object
Changes in version 1.14.0:
o Added functions for calculating Shapley indices: shapleyPermEx()
and shapleyPermRand()
o Added function for calculating first order Sobol' indices via
smoothing technique: sobolSmthSpl()