Mixture Hidden Markov Models for Social Sequence Data and Other
Multivariate, Multichannel Categorical Time Series
Designed for fitting hidden (latent) Markov models and mixture
hidden Markov models for social sequence data and other categorical time series.
Also some more restricted versions of these type of models are available: Markov
models, mixture Markov models, and latent class models. The package supports
models for one or multiple subjects with one or multiple parallel sequences
(channels). External covariates can be added to explain cluster membership in
mixture models. The package provides functions for evaluating and comparing
models, as well as functions for visualizing of multichannel sequence data and
hidden Markov models. Models are estimated using maximum likelihood via the EM
algorithm and/or direct numerical maximization with analytical gradients. All
main algorithms are written in C++ with support for parallel computation.
Documentation is available via several vignettes in this page, and the
paper by Helske and Helske (2019, ).
News
seqHMM 1.0.12 (Release date: 2019-04-11)
Fixed an OpenMP sharing issue due to upcoming GCC9.
seqHMM 1.0.11 (Release date: 2019-04-09)
All functions relying on C++ are now slightly faster due to
more efficient data transformations from factors to integers.
Fixed an OpenMP sharing issue due to upcoming GCC9.
seqHMM 1.0.10 (Release date: 2019-01-25)
Fixed iteration counter for EM algorithm, now maxeval=1 actually does
one iteration and not two. Similarly fixed the the corresponding printing of
intermediate results.
build_hmm and build_hmm now ignore n_states argument if any of the initial values
are provided.
Updated some formulas in the algorithms vignette to correspond to current forward-backward variant.
Updated citation info due to the publication in JSS.
seqHMM 1.0.9 (Release date: 2018-11-06)
Fixed a bug in backward and posterior probabilities of multivariate models stemming
from changes in version 1.0.7, causing the incorrect scaling factors. Note that issue
was only in forward_backward and posterior_probs functions with log_scale = FALSE,
internally the package still used correct implementations when calling fit_model.
Thanks for Silvia Bacci for noticing the issue.
Fixed OpenMP flags in Makevars.
seqHMM 1.0.8-1 (Release date: 2018-05-03)
Fixed encoding issues in references.
Updated author affiliations.
seqHMM 1.0.8 (Release date: 2017-11-07)
Fixed a bug in forward_backward function, which did not provide correct
scaling factors for last time point and thus the backward variables weren't F
scaled correctly.
Related to above bug, posterior_probs now provides proper probabilities
between 0 and 1.
Fixed a bug in ssp; caused an error for sortv = "hidden.paths" when hidden
paths were not provided even though x was an hmm object.
Changed argument withlegend to with.legend due to changes in the TraMineR package.
Related to the TraMineR update, fixed warnings given by ssp functions.
Fixing model plots after an update in the igraph package.
seqHMM 1.0.7 (Release date: 2017-04-04)
Added supplementary vignettes for visualization, estimation, and theoretical
background.
Corrected a "feature" of the scaled forward-backward algorithm which caused
potential numerical issues in backward probabilities. More specifically,
previously the scaled backward variable at time t was scaled with the scaling
factors c_[t+1] of the forward variable, instead of c_t. In typical
applications this did not cause any problems, but some models which previously
caused problems can work now without resorting to the log-space algorithm.
Scaling factors used in forward-backward algorithms are now stored as 1/c_t,
where c_t are the scaling factors from older versions of seqHMM.
The build_mm function now automatically estimates model parameters from the
observed initial state distribution and transition counts.
Added automatic starting values for build_hmm, build_mhmm, build_mmm, and
build_lcm.
Bug fixes:
Fixed a bug from version 1.0.6 where the EM algorithm with restarts did not
work as efficiently as intended due to changes in initializing emission matrices.
More specifically, the previous version used Rcpp sugar which does not use deep
copies in case of arma::cubes, which resulted the subsequent EM runs to depend
on the first optimum (affecting especially the identification of non-zero
emission probabilities).
Fixed a bug in the ssp function where tlim was taken account of after sorting
sequences and computing hidden paths (now cases are chosen before other actions).
seqHMM 1.0.6 (Release date: 2016-08-01)
Argument diag_c in simulate_transition_probs is now used also in cases where
left_right = FALSE.
Adjusted reltol and maxeval values for EM algorithm. Now reltol is 1e-10
(previously 1e-12), and the reltol and maxeval values for restarts are by
default taken from the initial EM algorithm (previosly reltol was 1e-8 and
maxeval = 100 for restarts).
Fixed hidden states labels for ssp functions (previously always used the
default values).
Bug fixes:
The ssplot function assigned wrong colors for hidden states in cases where
the state names were not alphabetically ordered. The performance of the
function was also improved by removing extra calls to seqdef.
Changing the missing.color argument did not work in legends of ssp, ssplot, and
mssplot.
The mssplot function now works with unique hidden state names (problem
occured e.g. with latent class models).
The mssplot function with sortv = "mds.hidden" produced strange errors when
plotting clusters with one hidden state. Now automatically uses "mds.obs" in
such cases.
seqHMM 1.0.5 (Release date: 2016-02-24)
Bug fixes:
The mssplot function now uses hidden paths instead of posterior probabilities to determine the most probable cluster for each subject (previous solution caused errors when posterior probabilities suggested a different cluster than hidden paths).
In mssplot, removed a misplaced tlim which caused a warning when plotting state distributions of hidden paths.
The gridplot function now uses with.missing.legend also with combined legends.
seqHMM 1.0.4 (Release date: 2016-01-14)
Added examples for build_mmm.
Added more space for main titles in plot.mhmm.
Improved documentation.
Added tlim in ssp functions.
Bug Fixes:
Corrected mc_to_sc for single state models due to dimension dropping.
Corrected a bug in plot.hmm when layout = "vertical".
Fixed legend layout in plot.hmm.
Wrong nobs and df attributes in mc_to_sc.
Wrong number of sequences to ssp titles when tlim is used.
EM with HMM using log-space produced error due to missing element in output of EM.
seqHMM 1.0.3-1 (Release date: 2015-12-29)
As requested by CRAN, changed donttest example of interactive plotting to conditional block depending
on whether session is interactive or not.
seqHMM 1.0.3 (Release date: 2015-12-23)
Corrected dependency on R 3.2.0 due to lengths function.
Bug Fixes:
Corrected a bug which caused fit_model to stop if restarted EM failed.
vcov.mhmm produced errors in valgrind, corrected issue by replacing Armadillo's inv_sympd function with inv.
Corrected a bug relating to colorpalette in mc_to_sc function.
Performance improvements:
Slight performance improvement in all functions by tweaking the usage of armadillo constructors.