NEWS

---

mixvlmc 0.2.1.9000

Breaking changes

• draw() has now a second parameter format with no default. This may break codes that used the fact that the control parameter was previously the second one.
• the digits parameter of draw.covlmc() has been removed and replaced by a similar parameter in draw_control()
• defaults for draw.covlmc() have changed (issue #66):
  • p_value is now FALSE by default
  • when p_value is TRUE the p-values are represented between separators given by the open_p_value and close_p_value members of the control list. Defaults are < and >
  • the intercept is separated from other coefficients by the characters specified in the intercept_sep member of the control list. Defaults to " & "
  • the characters used to separate levels from models is now " ~ "
  • temporal blocks are now separated by default by the time_sep member of the control list. The default value is set to " | " and the time_sep parameter has been removed from draw.covlmc()
  • the new parameter constant_as_prob is set to TRUE which modifies also the default display
• the interface of draw_control() has been significantly changed (issue #66) which should break most customisation code at the level of characters used by draw() to display a model

New features

Discrete time series class (dts)

In order to ease the introduction of multiple time series support, a single discrete time series can now be represented by the dts class, via the dts() function (see issue #76). All functions that use discrete time series now accept objects of this class in addition to simple vectors of a supported type (integer, factor, character and logical). This applies to model estimation functions such as vlmc() or covlmc(), to model selection functions (e.g. tune_vlmc) but also to functions that use new data such loglikelihood() and predict.vlmc().

Model representation (with draw())

A major change of draw() is the support of multiple output formats. This is done via a format parameter. It supports currently:

• the format of previous versions of mixvlmc with format="text". This is the default text based representation.
• a new LaTeX export with format="latex" based on the LaTeX package forest (https://ctan.org/pkg/forest) as per issue #66. This can be used to include (CO)VLMC models in LaTeX document, for instance when using rmarkdown or quarto.

In addition, text based model representation has been improved (as per issue #66) as follows:

• a new global option mixvlmc.charset can be used to select the characters used by draw() when format="text" between pure ASCII and UTF-8
• when possible, the text based representation will default to UTF-8 symbols to provide a cleaner display of the context tree
• full customisation of the character set has moved from draw_control() to charset_ascii() and charset_utf8()
• the draw.covlmc() function uses arguably better default parameters (described above)
• when a logistic model does not use the covariates, it is now represented by draw.covlmc() as a VLMC node, that is with the probability distribution on the state space that it represents

mixvlmc 0.2.1 (2023-10-31)

This release has no visible changes and fixes internal issues:

• missing package in Suggests
• double definition of a C++ internal function

mixvlmc 0.2.0 (2023-10-28)

Breaking changes

• the mode parameter of cutoff.vlmc() has been renamed to scale
• the type parameter has been removed from contexts() variants and the default result format has been modified significantly
• the counts parameter from contexts.vlmc() and contexts.covlmc() has been replaced by a logical parameter named local
• contexts() sequences are now reported by default in temporal order

New features

C++ backend

The main major new feature of this version is the inclusion of a C++ implementation of context tree and VLMC construction. A new option mixvlmc.backend can be used to switch globally from the original "R" implementation to the new "C++" back end. A new backend parameter has been added to ctx_tree(), vlmc() and tune_vlmc() to enable local back end selection.

The C++ implementation is significantly faster than the R implementation, at least by a factor 10. While it has been thoroughly tested, it is still considered experimental notably because it does not apply to COVLMC (setting the global option to "C++" has not effect on COVLMC model construction). For context trees and VLMC, results should not depend on the back end, at least within numerical precision. The only notable difference is the ordering of the contexts which differs between back ends: in a call to contexts(), the first context for the R back end will generally not be the first context for the C++ back end.

Other new features

• contexts() can now report the positions of each context in the original time series
• nodes of context trees can be extracted individually as ctx_node objects using the find_sequence() function. A collection of new functions can be used to manipulate the nodes and gain fine grain information on the corresponding sequences (issue #50).
• by default, contexts() reports now contexts as a list of ctx_node objects
• new functions predict.vlmc() and predict.covlmc() can be used to make one step ahead predictions of a time series based on a (CO)VLMC model (issue #46). Those function are documented in a new vignette (vignette("prediction"))
• the log likelihood calculations performed by logLik.vlmc(), logLik.covlmc() loglikelihood() and loglikelihood.covlmc() have been revised, expanded to include three possible definitions of the likelihood function, and documented in a new vignette (vignette("likelihood"))
• tune_vlmc() and tune_covlmc() can be used with the different likelihood function definitions
• results of tune_vlmc() and tune_covlmc() can be plotted using base R graphics or ggplot2 (issue #36)
• tune_covlmc() can trim the best model (and the initial one) if asked to
• trim.covlmc() implements simple trimming for VGAM based objects (issue #48)
• cutoff() uses a new tolerance parameter to avoid reporting cut off values that are almost identical due to numerical imprecision
• simulate.vlmc() implements a user specified burn in period (issue #40)
• simulate.vlmc() and simulate.covlmc() now handle the random generator state as does stats::simulate() (issue #56)

Minor improvements and bug fixes

• Improved the test coverage (issue #54)
• Fixed several bugs in internal sampling methods that made simulate.covlmc() unreliable for state spaces with three or more states
• Fixed a bug in simulate.covlmc() that occurred in contexts with a longer self memory compared to their covariate memory
• Fixed several bugs in metrics.vlmc() and align the results with the ones obtained by using direct calculation on the results of predict.vlmc()
• Fixed several bugs related to degenerate VLMC (issue #55)
• Fixed a bug in the likelihood calculation for internal context nodes
• Improved type consistency by reporting counts as integers in contexts() results
• Fixed documentation: all models can be adjusted on time series of logical values
• Synchronized ctx_tree() documentation and its default value
• Synchronized contexts() documentation and its default value
• Added more trimming for COVLMC models in trim.covlmc()

mixvlmc 0.1.1 (2023-07-10)

• Added missing documentation about the return values of simulate.vlmc() and simulate.covlmc()

mixvlmc 0.1.0

• Added a NEWS.md file to track changes to the package.