disclapmix: Discrete Laplace mixture inference using the EM algorithm

In mikldk/disclapmix: Discrete Laplace Mixture Inference using the EM Algorithm

Discrete Laplace mixture inference using the EM algorithm

Description

disclapmix makes inference in a mixture of Discrete Laplace distributions using the EM algorithm. After the EM algorithm has converged, the centers are moved if the marginal likelihood increases by doing so. And then the EM algorithm is run again. This continues until the centers are not moved.

Usage

disclapmix(
  x,
  clusters,
  init_y = NULL,
  iterations = 100L,
  eps = 0.001,
  verbose = 0L,
  glm_method = "internal_coef",
  glm_control_maxit = 50L,
  glm_control_eps = 1e-06,
  init_y_method = "pam",
  init_v = NULL,
  ret_x = FALSE,
  ...
)

Arguments

x	Dataset.
clusters	The number of clusters/components to fit the model for.
init_y	Initial central haplotypes, if NULL, these will be estimated as described under the init_y_method argument.
iterations	Maximum number of iterations in the EM-algorithm.
eps	Convergence stop criteria in the EM algorithm which is compared to \frac{\max \{ v_{new} - v_{old} \}}{\max \{ v_{old} \}}, where v is a matrix of each observation's probability of belonging to a certain center.
verbose	from 0 to 2 (both including): 0 for silent, 2 for extra verbose.
glm_method	internal_coef, internal_dev or glm.fit. Please see details.
glm_control_maxit	Integer giving the maximal number of IWLS iterations.
glm_control_eps	Positive convergence tolerance epsilon; the iterations converge when |x - x_{old}|/(|x| + 0.1) < epsilon, where x = beta_correction for internal_coef and x = deviance otherwise.
init_y_method	Which cluster method to use for finding initial central haplotypes, y: pam (recommended), clara or hclust. Ignored if init_y is supplied.
init_v	Matrix with 'nrow(x)' rows and 'clusters' columns specifying initial posterior probabilities to get EM started, if none specified, then 'matrix(1/clusters, nrow = nrow(x), ncol = clusters)' is used.
ret_x	Return data 'x'
...	Used to detect obsolete usage (when using parameters centers, use.parallel, calculate.logLs or plots.prefix).

Details

glm_method: internal_coef is the fastest as it uses the relative changes in the coefficients as a stopping criterium, hence it does not need to compute the deviance until the very end. In normal situations, it would not be a problem to use this method. internal_dev is the reasonably fast method that uses the deviance as a stopping criterium (like glm.fit). glm.fit to use the traditional glm.fit IWLS implementation and is slow compared to the other two methods.

init_y_method: For init_y_method = 'clara', the sampling parameters are: samples = 100, sampsize = min(ceiling(nrow(x)/2), 100 + 2*clusters) and the random number generator in R is used.

Value

A disclapmixfit object:

list("glm_method")

The supplied GLM method.

list("init_y")

The supplied initial central haplotypes, init_y.

list("init_y_method")

The supplied method for choosing initial central haplotypes (only used if init_y is NULL).

list("converged")

Whether the estimation converged or not.

list("x")

Dataset used to fit the model if 'ret_x' is 'TRUE', else 'NULL'.

list("y")

The central haplotypes, y.

list("tau")

The prior probabilities of belonging to a cluster, tau.

list("v_matrix")

The matrix v of each observation's probability of belonging to a certain cluster. The rows are in the same order as the observations in x used to generate this fit.

list("disclap_parameters")

A matrix with the estimated dicrete Laplace parameters.

list("glm_coef")

The coefficients from the last GLM fit (used to calculate disclap_parameters).

list("model_observations")

Number of observations.

list("model_parameters")

Number of parameters in the model.

list("iterations")

Number of iterations performed in total (including moving centers and re-estimating using the EM algorithm).

list("logL_full")

Full log likelihood of the final model.

list("logL_marginal")

Marginal log likelihood of the final model.

list("BIC_full")

BIC based on the full log likelihood of the final model.

list("BIC_marginal")

BIC based on the marginal log likelihood of the final model.

list("v_gain_iterations")

The gain \frac{\max \{ v_{new} - v_{old} \}}{\max \{ v_{old} \}}, where v is vic_matrix mentioned above, during the iterations.

list("tau_iterations")

The prior probability of belonging to the centers during the iterations.

list("logL_full_iterations")

Full log likelihood of the models during the iterations (only calculated when verbose = 2L).

list("logL_marginal_iterations")

Marginal log likelihood of the models during the iterations (only calculated when verbose = 2L).

list("BIC_full_iterations")

BIC based on full log likelihood of the models during the iterations (only calculated when verbose = 2L).

list("BIC_marginal_iterations")

BIC based on marginal log likelihood of the models during the iterations (only calculated when verbose = 2L).

See Also

disclapmix-package disclapmix disclapmixfit predict.disclapmixfit print.disclapmixfit summary.disclapmixfit simulate.disclapmixfit clusterdist clusterprob glm.fit disclap pam clara

Examples

# Generate sample database
db <- matrix(disclap::rdisclap(1000, 0.3), nrow = 250, ncol = 4)

# Add location parameters
db <- sapply(1:ncol(db), function(i) as.integer(db[, i]+13+i))

head(db)

fit1 <- disclapmix(db, clusters = 1L, verbose = 1L, glm_method = "glm.fit")
fit1$disclap_parameters
fit1$y

fit1b <- disclapmix(db, clusters = 1L, verbose = 1L, glm_method = "internal_coef")
fit1b$disclap_parameters
fit1b$y

max(abs(fit1$disclap_parameters - fit1b$disclap_parameters))

# Generate another type of database
db2 <- matrix(disclap::rdisclap(2000, 0.1), nrow = 500, ncol = 4)
db2 <- sapply(1:ncol(db2), function(i) as.integer(db2[, i]+14+i))
fit2 <- disclapmix(rbind(db, db2), clusters = 2L, verbose = 1L)
fit2$disclap_parameters
fit2$y
fit2$tau

mikldk/disclapmix documentation built on Aug. 22, 2023, 10:56 a.m.