- An error in compute_mallows_loglik when the number of clusters is more than one has been corrected. Thanks to Marta Crispino.
- Skipping a unit test which failed on CRAN's M1 Mac builder.
- For initialization of latent ranks when using pairwise preference data, all topological sorts are now generated in random order.
- The SMC function now check for consistency with previous latent ranks for existing users also when data arrive in the form of pairwise preferences.
- A function compute_exact_partition_function() is now added, which returns the logarithm of the exact partition function for Cayley, Hamming, and Kendall distance.
- Fixed a bug in the Ulam distance. Thanks for Marta Crispino for discovering it.
- Fixed a bug in SMC algorithm for pairwise preference data, where the proposal distribution incorrectly was assumed to be uniform.
- It is now possible to report progress of MCMC more flexibly using compute_mallows() or compute_mallows_mixtures(). The old argument "verbose" which by default reported every 1000'th iteration has been replaced by an argument "progress_report" which can be set by calling set_progress_report(). The latter allows setting the interval between reports. This is particularly useful for big data, where running 1000 iterations may take very long time.
- Fixed a bug which caused inconsistent partial rank data to be retained from previous timepoints when existing users update their preferences.
- Arguments random and random_limit to setup_rank_data() have been removed. A new argument max_topological_sorts has been added instead, which captures all previous use cases, but also allows the user to specify the number of topological sorts to generate. This makes it useful also with a relatively large number of items, while it previously would be computationally unfeasible for anything more than 8-9 items.
- Argument shuffle_unranked to setup_rank_data() has been removed. If there are unranked items they will now always be shuffled. For reproducibility, set the random number seed.
- SMC Mallows with pairwise preference data now allows different initially values for the augmented rankings across the particles. This is obtained by generating (a subset of) all topological sorts consistent with the transitive closure for the user, and sampling from these. Can be set with the max_topological_sorts argument to set_smc_options().
- Fixed gcc-UBSAN issue happening when compute_mallows_sequentially() is run without user IDs specified.
- The SMC method update_mallows() now supports pairwise preferences, both new users providing pairwise preferences and existing users updating their preferences.
- Acceptance ratios are now tracked both in the Metropolis-Hastings algorithm used by compute_mallows() and in the move step inside the sequential Monte Carlo algorithm used by update_mallows() and compute_mallows_sequentially(). Use the function get_acceptance_ratios() to access them.
- BREAKING CHANGE: Burnin now has to be explicitly set using 'burnin(model) <- value' if it is not already set in compute_options. This alleviates the need for a 'burnin' argument in the functions for assessing the posterior distribution and it abstracts away the implementation from the user. See '?burnin' and '?burnin<-' for details.
- The swap proposal defined in Crispino et al., Annals of Applied Statistics (2019) is now an option for proposing the modal ranking rho. It can be defined by setting rho_proposal="swap" in set_compute_options(). The leap-and- shift distribution is still the default.
- Fixed a bug in heat_plot() when the model has been estimated with rho_thinning > 1, causing the probabilities to be unnormalized. Issue #381. Thanks to Marta Crispino for discovering the bug.
- Added stratified, systematic, and residual resampling to the sequential Monte Carlo algorithm. These distributions should in general be preferred to multinomial resampling, which was the only available option until now.
- The move step of the SMC algorithm now allows a user-defined lag for the sampling of latent ranks, specified in the "latent_sampling_lag" argument to set_smc_options().
- Prior for precision parameter alpha is now a gamma distribution. Until now an exponential distribution has been assumed. Since the exponential is a special case of the gamma with shape parameter equal to 1 (the default), this is not a breaking change. However, it adds flexibility when it comes to specifying the prior.
- setup_rank_data() now accepts a single vector of rankings, silently converting a vector to matrix with a single row.
- Sequential Monte Carlo algorithm can now start from a sample from the prior distribution, see the sample_prior() function for an example.
- Added support for parallelism under-the-hood with oneTBB.
- Edits to C++ code fixing memory leaks.
- Edits to unit tests which caused issues on CRAN.
- Large refactoring with several breaking changes. See vignettes and documentation for details.
- Bug in plot.BayesMallows for posterior distribution with 'parameter = "rho"' has been fixed. Thanks to Lorenzo Zuccato for points out the issue. (#342)
- Argument obs_freq to internal function rmallows() is removed, as it is not being used. Thanks to Lorenzo Zuccato for pointing this our (#337).
- Argument save_clus to compute_mallows() has been removed, as it was not used.
- compute_mallows() now supports parallel chains, by providing a 'cl' argument. See vignette "MCMC with Parallel Chains" for a tutorial.
- Documentation of functions are now grouped in families.
- lik_db_mix() is now deprecated in favor of get_mallows_loglik()
- Unusued argument removed from internal function augment_pairwise(). Thanks to Lorenzo Zuccato for making us aware of this (#313).
- Bug fix: psi argument to compute_mallows() and compute_mallows_mixtures(), specifying the concentration parameter of the Dirichlet prior, is now forwarded to the underlying run_mcmc() function. Previously, this argument has had no effect, and the default psi=10 has been used regardless of the input. Thanks to Lorenzo Zuccato for discovering this bug.
- SMC functions now accept exact partition functions where these are available.
- Removed SMC functions deprecated on version 1.2.0 (#301)
- Website deployed at https://ocbe-uio.github.io/BayesMallows.
- Reordering of authors, so Waldir Leoncio appears second in the list.
- Fixed LTO compilation notes on CRAN.
- Fixed package documentation issue on CRAN.
- Added heat_plot() function (#255)
- Replaced deprecated ggplot2::aes_ function with ggplot2::aes.
- Refactoring of SMC functions (#257)
- Improved validation and documentation of SMC post-processing functions (#262)
- Added
plot.SMCMallows()method - Changed default values and argument order on several SMC functions (see PR #269)
- Modifications to internal C++ code to avoid CRAN NOTEs.
- PerMallows package has been removed from Imports because it is at risk of being removed from CRAN. This means that for Ulam distance with more than 95 items, the user will have to compute an importance sampling estimate.
- Refactoring of data augmentation function for SMC Mallows.
- Improved documentation of
sample_dataset
- Fixed a bug which caused assess_convergence() to fail with 'parameter = "cluster_probs"'.
- Fixed a bug in smc_mallows_new_users_partial() and smc_mallows_new_users_partial_alpha_fixed().
- metropolis_hastings_aug_ranking_pseudo() has been deprecated. Please use metropolis_hastings_aug_ranking() instead, with pseudo=TRUE.
- smc_mallows_new_users_partial_alpha_fixed(), smc_mallows_new_users_complete(), and smc_mallows_new_users_partial() have been deprecated. Please use smc_mallows_new_users() instead, and set the type= argument to "complete", "partial", or "partial_alpha_fixed".
- smc_mallows_new_item_rank_alpha_fixed() has been deprecated. Please use smc_mallows_new_item_rank() instead, with argument alpha_fixed=TRUE.
- Fixed unexpected behavior in leap-and-shift proposal distribution for SMC Mallows, causing the function to propose the current rank vector with nonzero probability.
- BayesMallows no longer depends on 'dplyr'.
- Quite extensive internal refactoring of C++ code.
- Function lik_db_mix has been renamed to get_mallows_loglik. lik_db_mix still exists as deprecated.
- When no initial rankings are provided, compute_mallows() and compute_mallows_mixtures() no use independent initial rho in each cluster. Previously a single initial rho was used for all cluster. This should potentially improve convergence, but will lead to different results when n_clusters>=2 for a given random number seed.
- Fixed an issue with stats::reshape causing an error on R-oldrel.
- Fixed an issue with checking the class of objects, where we now consistently use inherits().
- Internal C++ fixes to comply with CRAN checks.
- Fixed C++ errors leading to CRAN issues.
- Major update, introducing a whole new class of methods using sequential Monte Carlo. Also reducing the number of dependencies.
- This is a major update, with new functions for estimating the Bayesian Mallows model using sequential Monte Carlo. The methods are described in the vignette titled "SMC-Mallows Tutorial".
- Removed a large number of dependencies by converting to base R code. This will make the package easier to install across a range of systems, and less vulnerable to changes in other packages.
- Incorporates changes since 1.0.3, and also remove PLMIX from Imports.
- Fixed bug which caused plot_top_k to fail when plotting clusters.
- Improved the default value of rel_widths argument to plot_top_k.
- Wrote unit tests to check that the bugs don't appear again.
- Fixed bug which caused importance sampling to fail when running in parallel.
- Fixed issue with error message when trying to plot error probability when compute_mallows has not been set up to compute error probability.
- Increased number of unit tests.
- Fixed critical bug which caused results to be wrong with more than one mixture component in compute_mallows() and compute_mallows_mixtures(). Thanks to Anja Stein for discovering the bug.
- Function generate_initial_ranking() now has two additional options for generating random initial rankings. This can help with convergence problems, by allowing the MCMC algorithm to run from a range of different starting points.
- Fixes a bug in lik_db_mix and expected_dist, in which the scaling parameter used a different parametrization than the rest of the package. All functions in the package now use consistent parametrization of the Mallows model, as stated in the vignette.
- Function for computing likelihood added.
- Options for dealing with missing values added, and documentation now states how missing values are dealt with.
- Function rank_freq_distr added, for computing the frequency distribution of ranking patterns.
- Function rank_distance added, for computing the distance between rankings.
- Function expected_dist added for computing expectation of several metrics under the Mallows model.
- Function compute_consensus now includes an option for computing consensus of augmented ranks.
- Fixed bug in predict_top_k and plot_top_k when using aug_thinning > 1.
- Updated README and vignette.
- Updating a unit test to make sure BayesMallows is compatible with dplyr version 1.0.0.
- Improvement of plotting functions, as noted below.
- plot.BayesMallows and plot_elbow no longer print titles automatically.
- assess_convergence no longer prints legends for clusters, as the cluster number is essentially arbitrary.
- Added CITATION.
- Updated test of random number seed.
- Implements all fixes since version 0.3.1 below.
- Fixed typo on y-axis label of elbow plot.
- Fixed an issue which caused the cluster probabilities to differ across platforms, despite using the same seed. https://stackoverflow.com/questions/54822702
- Fixed a bug which caused
compute_mallowsnot to work (without giving any errors) whenrankingscontained missing values. - Fixed a bug which caused
compute_mallowsto fail whenpreferenceshad integer columns.
- Changed the name of
save_individual_cluster_probstosave_ind_clus, to save typing.
- Added a user prompt asking if the user really wants to save csv files, when
save_individual_cluster_probs = TRUEin compute_mallows. - Added
alpha_max, the truncation of the exponential prior foralpha, as a user option incompute_mallows.
- Added functionality for checking label switching. See
?label_switchingfor more info.
- The internal function
compute_importance_sampling_estimatehas been updated to avoid numerical overflow. Previously, importance sampling failed at below 200 items. Now it works way above 10,000 items.
- This is an update of some parts of the C++ code, to avoid failing the sanitizer checks clang-UBSAN and gcc-UBSAN.
- See all bullet points below, since 0.2.0.
generate_transitive_closure,generate_initial_ranking, andgenerate_constraintsnow are able to run in parallel.- Large changes to the underlying code base which should make it more maintainable but not affect the user.
estimate_partition_functionnow has an option to run in parallel, leading to significant speed-up.
- Implemented the Bernoulli error model. Set
error_model = "bernoulli"incompute_mallowsin order to use it. Examples will come later.
- Added parallelization option to
compute_mallows_mixturesand addedparallelto Suggests field.
- Deprecated functions
compute_cp_consensusandcompute_map_consensushave been removed. Usecompute_consensusinstead.
- Clusters are now
factorvariables sorted according to the cluster number. Hence, in plot legends, "Cluster 10" comes after "Cluster 9", rather than after "Cluster 1" which it used to do until now, because it was acharacter. plot.BayesMallowsno longer contains print statements which forces display of plots. Instead plots are returned from the function. Usingp <- plot(fit)hence does no longer display a plot, whereas usingplot(fit)without assigning it to an object, displays a plot. Until now the plot was always shown forrhoandalpha.
compute_mallowsandsample_mallowsnow support Ulam distance, with argumentmetric = "ulam".- Slimmed down the vignette significantly, in order to avoid clang-UBSAN error caused by running the vignette (which was then again caused by
Rcpp, cf. this issue). The long vignette is no longer needed in any case, since all the functions are well documented with executable examples.
- New release on CRAN, which contains all the updates in 0.1.1, described below.
Rankclusterpackage has been removed from dependencies.
- Fixed bug with Cayley distance. For this distance, the computational shortcut on p. 8 of Vitelli et al. (2018), JMLR, does not work. However, it was still used. Now, Cayley distance is always computed with complete rank vectors.
- Fixed bug in the default argument
leap_sizetocompute_mallows. It used to befloor(n_items / 5), which evaluates to zero whenn_items <= 4. Updated it tomax(1L, floor(n_items / 5)). - Added Hamming distance (
metric = "hamming") as an option tocompute_mallowsandsample_mallows.
- Updated
generate_initial_ranking,generate_transitive_closure, andsample_mallowsto avoid errors when packagetibbleversion 2.0.0 is released. This update is purely internal.
- Objects of class
BayesMallowsandBayesMallowsMixturesnow have default print functions, hence avoiding excessive amounts of informations printed to the console if the user happens to write the name of such an object and press Return. compute_mallows_mixturesno longer setsinclude_wcd = TRUEby default. The user can choose this argument.compute_mallowshas a new argumentsave_clus, which can be set toFALSEfor not saving cluster assignments.
assess_convergencenow automatically plots mixtures.compute_mallows_mixturesnow returns an object of classBayesMallowsMixtures.
assess_convergencenow adds prefix Assessor to plots whenparameter = "Rtilde".predict_top_kis now an exported function. Previously it was internal.
compute_posterior_intervalsnow has defaultparameter = "alpha". Until now, this argument has had no default.- Argument
typetoplot.BayesMallowsandassess_convergencehas been renamed toparameter, to be more consistent.
- Argument
save_augment_datatocompute_mallowshas been renamed tosave_aug. compute_mallowsfills in implied ranks when an assessor has only one missing rank. This avoids unnecessary augmentation in MCMC.generate_rankingandgenerate_orderingnow work with missing ranks.
Argument cluster_assignment_thinning to compute_mallows has been renamed to clus_thin.
Change the interface for computing consensus ranking. Now, CP and MAP consensus are both computed with the compute_consensus function, with argument type equal to either "CP" or "MAP".