Improve docs

R/compute_sequentially.R

@@ -1,5 +1,33 @@
 #' Compute the Bayesian Mallows model sequentially
 #'
+#' This function implements the nested sequential Monte Carlo (SMC2) algorithm
+#' for sequential learning of rank and preference data using the Bayesian
+#' Mallows model. The algorithm processes data sequentially over time,
+#' maintaining a particle approximation to the posterior distribution.
+#'
+#' @details
+#' The nested SMC2 algorithm consists of two levels of sequential Monte Carlo:
+#' 
+#' **Outer SMC (Parameter Level)**: Maintains particles representing samples
+#' from the posterior distribution of static parameters (alpha, rho, tau).
+#' Each particle contains its own set of parameter values.
+#' 
+#' **Inner SMC (Latent State Level)**: For each parameter particle, maintains
+#' multiple particle filters that track the evolution of latent rankings and
+#' cluster assignments over time. This nested structure allows the algorithm
+#' to handle the complex dependency structure between parameters and latent
+#' states.
+#' 
+#' At each timepoint, the algorithm:
+#' 1. Propagates each parameter particle forward using MCMC moves
+#' 2. For each parameter particle, runs multiple particle filters to sample
+#'    latent rankings and cluster assignments
+#' 3. Computes importance weights based on the likelihood of observed data
+#' 4. Resamples particles when effective sample size drops below threshold
+#' 5. Applies rejuvenation moves to maintain particle diversity
+#' 
+#' The nested structure is essential for maintaining proper uncertainty
+#' quantification in the joint parameter-latent state space.
 #'
 #' @param data A dataframe containing partial rankings or pairwise preferences.
 #'   If `data` contains complete or partial rankings, it must have the following
@@ -24,12 +52,93 @@
 #' }
 #'
 #' @param hyperparameters A list returned from [set_hyperparameters()].
-#' @param smc_options A list returned from [set_smc_options()]
+#' @param smc_options A list returned from [set_smc_options()]. Controls the
+#'   nested SMC2 algorithm parameters including number of parameter particles,
+#'   number of particle filters per parameter particle, and MCMC move parameters.
 #' @param topological_sorts A list returned from
 #'   [precompute_topological_sorts()]. Only used with preference data, and
-#'   defaults to `NULL`.
+#'   defaults to `NULL`. Contains precomputed topological sorts for efficient
+#'   sampling from constrained ranking spaces.
+#'
+#' @return An object of class BayesMallowsSMC2 containing posterior samples
+#'   and algorithm diagnostics.
+#'
+#' @references
+#' Sørensen, Ø., Stein, A., Netto, W. L., & Leslie, D. S. (2025). 
+#' Sequential Rank and Preference Learning with the Bayesian Mallows Model. 
+#' \emph{Bayesian Analysis}. DOI: 10.1214/25-BA1564.
+#'
+#' @seealso [set_hyperparameters()], [set_smc_options()], [precompute_topological_sorts()]
+#'
+#' @examples
+#' # Example with complete rankings
+#' library(BayesMallowsSMC2)
+#' 
+#' # Generate synthetic ranking data
+#' set.seed(123)
+#' n_items <- 5
+#' n_users <- 20
+#' n_timepoints <- 10
+#' 
+#' # Create synthetic data
+#' data <- expand.grid(
+#'   timepoint = 1:n_timepoints,
+#'   user = 1:n_users
+#' )
+#' 
+#' # Add random rankings for each item
+#' for(i in 1:n_items) {
+#'   data[[paste0("item", i)]] <- sample(1:n_items, nrow(data), replace = TRUE)
+#' }
+#' 
+#' # Set hyperparameters
+#' hyperparams <- set_hyperparameters(
+#'   n_items = n_items,
+#'   alpha_shape = 2,
+#'   alpha_rate = 1,
+#'   n_clusters = 2
+#' )
+#' 
+#' # Set SMC options
+#' smc_opts <- set_smc_options(
+#'   n_particles = 100,
+#'   n_particle_filters = 20,
+#'   metric = "kendall",
+#'   verbose = TRUE
+#' )
+#' 
+#' # Run sequential computation
+#' result <- compute_sequentially(
+#'   data = data,
+#'   hyperparameters = hyperparams,
+#'   smc_options = smc_opts
+#' )
+#' 
+#' # Example with pairwise preferences
+#' # First precompute topological sorts
+#' prefs_matrix <- matrix(c(1, 2, 2, 3, 3, 1), ncol = 2, byrow = TRUE)
+#' topo_sorts <- precompute_topological_sorts(
+#'   prefs = prefs_matrix,
+#'   n_items = 3,
+#'   save_frac = 0.1
+#' )
+#' 
+#' # Create preference data
+#' pref_data <- data.frame(
+#'   timepoint = c(1, 1, 2, 2),
+#'   user = c(1, 2, 1, 2),
+#'   top_item = c(1, 2, 3, 1),
+#'   bottom_item = c(2, 3, 1, 3)
+#' )
+#' 
+#' # Run with preferences
+#' result_prefs <- compute_sequentially(
+#'   data = pref_data,
+#'   hyperparameters = set_hyperparameters(n_items = 3),
+#'   smc_options = set_smc_options(n_particles = 50),
+#'   topological_sorts = topo_sorts
+#' )
 #'
-#' @return An object of class BayesMallowsSMC2.
 #' @export
 #'
 compute_sequentially <- function(

R/set_hyperparameters.R

@@ -1,12 +1,83 @@
-#' Set hyperparameters
+#' Set hyperparameters for Bayesian Mallows model
 #'
-#' @param n_items Integer defining the number of items.
-#' @param alpha_shape Shape of gamma prior for alpha.
-#' @param alpha_rate Rate of gamma prior for alpha.
-#' @param cluster_concentration Concentration parameter of Dirichlet distribution for cluster probabilities.
-#' @param n_clusters Integer defining the number of clusters.
+#' Configure prior distributions and model structure for the Bayesian Mallows
+#' model used in sequential estimation. This function sets hyperparameters
+#' for the precision parameters, modal rankings, and cluster structure.
+#'
+#' @details
+#' The Bayesian Mallows model assumes:
+#' 
+#' **Precision Parameters**: Each cluster k has a precision parameter alpha_k
+#' that controls the concentration around the modal ranking. Higher values
+#' indicate stronger agreement. The prior is Gamma(alpha_shape, alpha_rate).
+#' 
+#' **Modal Rankings**: Each cluster has a modal ranking rho_k that represents
+#' the "consensus" ranking for that cluster. These are sampled uniformly
+#' from the space of permutations.
+#' 
+#' **Cluster Probabilities**: The probability of assignment to each cluster
+#' follows a Dirichlet distribution with concentration parameter
+#' cluster_concentration.
+#' 
+#' **Cluster Structure**: The number of clusters must be specified a priori.
+#' Model selection can be performed by comparing models with different
+#' numbers of clusters.
+#'
+#' @param n_items Integer. Number of items being ranked. Must be provided
+#'   and determines the dimensionality of the ranking space.
+#' @param alpha_shape Numeric. Shape parameter of the Gamma prior distribution
+#'   for precision parameters alpha_k. Higher values concentrate the prior
+#'   around higher precision values. Default is 1 (exponential prior).
+#' @param alpha_rate Numeric. Rate parameter of the Gamma prior distribution
+#'   for precision parameters alpha_k. Higher values favor lower precision
+#'   (more dispersed rankings). Default is 0.5.
+#' @param cluster_concentration Numeric. Concentration parameter of the
+#'   Dirichlet prior for cluster assignment probabilities. Higher values
+#'   favor more equal cluster sizes, while lower values allow more unbalanced
+#'   clusters. Default is 10.
+#' @param n_clusters Integer. Number of mixture components (clusters) in the
+#'   model. Each cluster has its own modal ranking and precision parameter.
+#'   Default is 1 (single-cluster model).
+#'
+#' @return A list containing all hyperparameter values for use in
+#'   [compute_sequentially()].
+#'
+#' @references
+#' Sørensen, Ø., Stein, A., Netto, W. L., & Leslie, D. S. (2025). 
+#' Sequential Rank and Preference Learning with the Bayesian Mallows Model. 
+#' \emph{Bayesian Analysis}. DOI: 10.1214/25-BA1564.
+#'
+#' @seealso [compute_sequentially()], [set_smc_options()]
+#'
+#' @examples
+#' # Basic hyperparameters for 5 items, single cluster
+#' basic_hyper <- set_hyperparameters(n_items = 5)
+#' 
+#' # Multiple clusters with informative priors
+#' multi_cluster <- set_hyperparameters(
+#'   n_items = 10,
+#'   alpha_shape = 2,      # More concentrated precision prior
+#'   alpha_rate = 1,       # Moderate precision values
+#'   n_clusters = 3,       # Three mixture components
+#'   cluster_concentration = 5  # Allow unbalanced clusters
+#' )
+#' 
+#' # High-precision scenario (strong agreement expected)
+#' high_precision <- set_hyperparameters(
+#'   n_items = 8,
+#'   alpha_shape = 5,      # Strong prior for high precision
+#'   alpha_rate = 0.5,     # Favor high alpha values
+#'   n_clusters = 2
+#' )
+#' 
+#' # Low-precision scenario (weak agreement expected)
+#' low_precision <- set_hyperparameters(
+#'   n_items = 6,
+#'   alpha_shape = 1,      # Weak prior
+#'   alpha_rate = 2,       # Favor low alpha values
+#'   n_clusters = 1
+#' )
 #'
-#' @return A list
 #' @export
 #'
 set_hyperparameters <- function(

R/set_smc_options.R

@@ -1,26 +1,102 @@
-#' Set SMC options
-#'
-#' @param n_particles Number of particles
-#' @param n_particle_filters Initial number of particle filters for each
-#'   particle
-#' @param max_particle_filters Maximum number of particle filters.
-#' @param resampling_threshold Effective sample size threshold for resampling
-#' @param doubling_threshold Threshold for particle filter doubling. If the
-#'   acceptance rate of the rejuvenation step falls below this threshold, the
-#'   number of particle filters is doubled. Defaults to 0.2.
-#' @param max_rejuvenation_steps Maximum number of rejuvenation steps. If the
-#'   number of unique particles has not exceeded half the number of particles
-#'   after this many steps, the rejuvenation is still stopped.
-#' @param metric Metric
-#' @param resampler resampler
-#' @param latent_rank_proposal latent rank proposal
-#' @param verbose Boolean
-#' @param trace Logical specifying whether to save static parameters at each
-#'   timestep.
-#' @param trace_latent Logical specifying whether to sample and save one
-#'   complete set of latent rankings for each particle and each timepoint.
-#'
-#' @return A list
+#' Set SMC options for nested sequential Monte Carlo algorithm
+#'
+#' Configure parameters for the nested SMC2 algorithm used in sequential
+#' Bayesian Mallows model estimation. This function sets both outer-level
+#' (parameter particle) and inner-level (latent state particle filter)
+#' algorithm parameters.
+#'
+#' @details
+#' The nested SMC2 algorithm requires careful tuning of both levels:
+#' 
+#' **Outer SMC Level**: Controls the number of parameter particles and their
+#' rejuvenation through MCMC moves. More particles provide better approximation
+#' but increase computational cost.
+#' 
+#' **Inner SMC Level**: Each parameter particle maintains multiple particle
+#' filters for latent rankings. More filters improve latent state estimation
+#' but multiply computational cost by the number of parameter particles.
+#'
+#' @param n_particles Integer. Number of parameter particles in the outer SMC
+#'   sampler. Each particle represents a sample from the posterior distribution
+#'   of static parameters (alpha, rho, tau). Larger values improve posterior
+#'   approximation accuracy but increase computational cost linearly.
+#' @param n_particle_filters Integer. Initial number of particle filters
+#'   maintained by each parameter particle for sampling latent rankings and
+#'   cluster assignments. This creates the nested structure where each of the
+#'   `n_particles` parameter particles runs `n_particle_filters` inner particle
+#'   filters.
+#' @param max_particle_filters Integer. Maximum number of particle filters
+#'   allowed per parameter particle. The algorithm may double the number of
+#'   filters when rejuvenation acceptance rates are low, up to this limit.
+#' @param resampling_threshold Numeric. Effective sample size threshold for
+#'   triggering resampling of parameter particles. When ESS falls below this
+#'   value, multinomial resampling is performed. Default is `n_particles / 2`.
+#' @param doubling_threshold Numeric. Acceptance rate threshold for particle
+#'   filter doubling during rejuvenation. If the acceptance rate of MCMC
+#'   rejuvenation moves falls below this threshold, the number of particle
+#'   filters is doubled to improve mixing. Defaults to 0.2.
+#' @param max_rejuvenation_steps Integer. Maximum number of MCMC rejuvenation
+#'   steps applied to parameter particles. Rejuvenation stops early if the
+#'   number of unique particles exceeds half the total number of particles,
+#'   indicating sufficient diversity.
+#' @param metric Character. Distance metric for the Mallows model. Options
+#'   include "kendall", "cayley", "hamming", "footrule", "spearman", and "ulam".
+#'   Different metrics capture different aspects of ranking disagreement.
+#' @param resampler Character. Resampling algorithm for parameter particles.
+#'   Options include "multinomial", "residual", "stratified", and "systematic".
+#'   Systematic resampling often provides better performance.
+#' @param latent_rank_proposal Character. Proposal mechanism for sampling
+#'   latent rankings in the inner particle filters. Options include "uniform"
+#'   and other problem-specific proposals.
+#' @param verbose Logical. Whether to print algorithm progress and diagnostics
+#'   during execution. Useful for monitoring convergence and performance.
+#' @param trace Logical. Whether to save static parameter values (alpha, rho,
+#'   tau) from all particles at each timestep. Enables detailed posterior
+#'   analysis but increases memory usage.
+#' @param trace_latent Logical. Whether to sample and save one complete set
+#'   of latent rankings for each parameter particle at each timepoint. Provides
+#'   full posterior samples of latent states but significantly increases memory
+#'   requirements.
+#'
+#' @return A list containing all SMC2 algorithm parameters for use in
+#'   [compute_sequentially()].
+#'
+#' @references
+#' Sørensen, Ø., Stein, A., Netto, W. L., & Leslie, D. S. (2025). 
+#' Sequential Rank and Preference Learning with the Bayesian Mallows Model. 
+#' \emph{Bayesian Analysis}. DOI: 10.1214/25-BA1564.
+#'
+#' @seealso [compute_sequentially()], [set_hyperparameters()]
+#'
+#' @examples
+#' # Basic SMC options for small problems
+#' basic_opts <- set_smc_options(
+#'   n_particles = 100,
+#'   n_particle_filters = 20,
+#'   metric = "kendall"
+#' )
+#' 
+#' # High-precision options for larger problems
+#' precise_opts <- set_smc_options(
+#'   n_particles = 1000,
+#'   n_particle_filters = 100,
+#'   max_particle_filters = 500,
+#'   resampling_threshold = 500,
+#'   metric = "footrule",
+#'   resampler = "systematic",
+#'   verbose = TRUE,
+#'   trace = TRUE
+#' )
+#' 
+#' # Memory-efficient options
+#' efficient_opts <- set_smc_options(
+#'   n_particles = 200,
+#'   n_particle_filters = 30,
+#'   trace = FALSE,
+#'   trace_latent = FALSE,
+#'   verbose = FALSE
+#' )
+#'
 #' @export
 #'
 set_smc_options <- function(