covdepGE: Covariate Dependent Graph Estimation

In JacobHelwig/covdepGE: Covariate Dependent Graph Estimation

Covariate Dependent Graph Estimation

Description

Model the conditional dependence structure of X as a function of Z as described in (1)

Usage

covdepGE(
  X,
  Z = NULL,
  hp_method = "hybrid",
  ssq = NULL,
  sbsq = NULL,
  pip = NULL,
  nssq = 5,
  nsbsq = 5,
  npip = 5,
  ssq_mult = 1.5,
  ssq_lower = 1e-05,
  snr_upper = 25,
  sbsq_lower = 1e-05,
  pip_lower = 1e-05,
  pip_upper = NULL,
  tau = NULL,
  norm = 2,
  center_X = TRUE,
  scale_Z = TRUE,
  alpha_tol = 1e-05,
  max_iter_grid = 10,
  max_iter = 100,
  edge_threshold = 0.5,
  sym_method = "mean",
  parallel = FALSE,
  num_workers = NULL,
  prog_bar = TRUE
)

Arguments

X	n \times p numeric matrix; data matrix. For best results, n should be greater than p
Z	NULL OR n \times q numeric matrix; extraneous covariates. If NULL, Z will be treated as constant for all observations, i.e.: Z <- rep(0, nrow(X)) If Z is constant, the estimated graph will be homogeneous throughout the data. NULL by default
hp_method	character in c("grid_search","model_average","hybrid"); method for selecting hyperparameters from the the hyperparameter grid. The grid will be generated as the Cartesian product of ssq, sbsq, and pip. Fix X_j, the j-th column of X, as the response; then, the hyperparameters will be selected as follows: If "grid_search", the point in the hyperparameter grid that maximizes the total ELBO summed across all n regressions will be selected If "model_average", then all posterior quantities will be an average of the variational estimates resulting from the model fit for each point in the hyperparameter grid. The unnormalized averaging weights for each of the n regressions are the exponentiated ELBO If "hybrid", then models will be averaged over pip as in "model_average", with \sigma^2 and \sigma_\beta^2 chosen for each \pi in pip by maximizing the total ELBO over the grid defined by the Cartesian product of ssq and sbsq as in "grid_search" "hybrid" by default
ssq	NULL OR numeric vector with positive entries; candidate values of the hyperparameter \sigma^2 (prior residual variance). If NULL, ssq will be generated for each variable X_j fixed as the response as: ssq <- seq(ssq_lower, ssq_upper, length.out = nssq) NULL by default
sbsq	NULL OR numeric vector with positive entries; candidate values of the hyperparameter \sigma_\beta^2 (prior slab variance). If NULL, sbsq will be generated for each variable X_j fixed as the response as: sbsq <- seq(sbsq_lower, sbsq_upper, length.out = nsbsq) NULL by default
pip	NULL OR numeric vector with entries in (0, 1); candidate values of the hyperparameter \pi (prior inclusion probability). If NULL, pip will be generated for each variable X_j fixed as the response as: pip <- seq(pip_lower, pi_upper, length.out = npip) NULL by default
nssq	positive integer; number of points to generate for ssq if ssq is NULL. 5 by default
nsbsq	positive integer; number of points to generate for sbsq if sbsq is NULL. 5 by default
npip	positive integer; number of points to generate for pip if pip is NULL. 5 by default
ssq_mult	positive numeric; if ssq is NULL, then for each variable X_j fixed as the response: ssq_upper <- ssq_mult * stats::var(X_j) Then, ssq_upper will be the greatest value in ssq for variable X_j. 1.5 by default
ssq_lower	positive numeric; if ssq is NULL, then ssq_lower will be the least value in ssq. 1e-5 by default
snr_upper	positive numeric; upper bound on the signal-to-noise ratio. If sbsq is NULL, then for each variable X_j fixed as the response: s2_sum <- sum(apply(X, 2, stats::var)) sbsq_upper <- snr_upper / (pip_upper * s2_sum) Then, sbsq_upper will be the greatest value in sbsq. 25 by default
sbsq_lower	positive numeric; if sbsq is NULL, then sbsq_lower will be the least value in sbsq. 1e-5 by default
pip_lower	numeric in (0, 1); if pip is NULL, then pip_lower will be the least value in pip. 1e-5 by default
pip_upper	NULL OR numeric in (0, 1); if pip is NULL, then pip_upper will be the greatest value in pip. If sbsq is NULL, pip_upper will be used to calculate sbsq_upper. If NULL, pip_upper will be calculated for each variable X_j fixed as the response as: lasso <- glmnet::cv.glmnet(X, X_j) non0 <- sum(glmnet::coef.glmnet(lasso, s = "lambda.1se")[-1] != 0) non0 <- min(max(non0, 1), p - 1) pip_upper <- non0 / p NULL by default
tau	NULL OR positive numeric OR numeric vector of length n with positive entries; bandwidth parameter. Greater values allow for more information to be shared between observations. Allows for global or observation-specific specification. If NULL, use 2-step KDE methodology as described in (2) to calculate observation-specific bandwidths. NULL by default
norm	numeric in [1, \infty]; norm to use when calculating weights. Inf results in infinity norm. 2 by default
center_X	logical; if TRUE, center X column-wise to mean 0. TRUE by default
scale_Z	logical; if TRUE, center and scale Z column-wise to mean 0, standard deviation 1 prior to calculating the weights. TRUE by default
alpha_tol	positive numeric; end CAVI when the Frobenius norm of the change in the alpha matrix is within alpha_tol. 1e-5 by default
max_iter_grid	positive integer; if tolerance criteria has not been met by max_iter_grid iterations during grid search, end CAVI. After grid search has completed, CAVI is performed with the final hyperparameters selected by grid search for at most max_iter iterations. Does not apply to hp_method = "model_average". 10 by default
max_iter	positive integer; if tolerance criteria has not been met by max_iter iterations, end CAVI. 100 by default
edge_threshold	numeric in (0, 1); a graph for each observation will be constructed by including an edge between variable i and variable j if, and only if, the (i, j) entry of the symmetrized posterior inclusion probability matrix corresponding to the observation is greater than edge_threshold. 0.5 by default
sym_method	character in c("mean","max","min"); to symmetrize the posterior inclusion probability matrix for each observation, the (i, j) and (j, i) entries will be post-processed as sym_method applied to the (i, j) and (j, i) entries. "mean" by default
parallel	logical; if TRUE, hyperparameter selection and CAVI for each of the p variables will be performed in parallel using foreach. Parallel backend may be registered prior to making a call to covdepGE. If no active parallel backend can be detected, then parallel backend will be automatically registered using: doParallel::registerDoParallel(num_workers) FALSE by default
num_workers	NULL OR positive integer less than or equal to parallel::detectCores(); argument to doParallel::registerDoParallel if parallel = TRUE and no parallel backend is detected. If NULL, then: num_workers <- floor(parallel::detectCores() / 2) NULL by default
prog_bar	logical; if TRUE, then a progress bar will be displayed denoting the number of remaining variables to fix as the response and perform CAVI. If parallel, no progress bar will be displayed. TRUE by default

Value

Returns object of class covdepGE with the following values:

graphs	list with the following values: graphs: list of n numeric matrices of dimension p \times p; the l-th matrix is the adjacency matrix for the l-th observation unique_graphs: list; the l-th element is a list containing the l-th unique graph and the indices of the observation(s) corresponding to this graph inclusion_probs_sym: list of n numeric matrices of dimension p \times p; the l-th matrix is the symmetrized posterior inclusion probability matrix for the l-th observation inclusion_probs_asym: list of n numeric matrices of dimension p \times p; the l-th matrix is the posterior inclusion probability matrix for the l-th observation prior to symmetrization
variational_params	list with the following values: alpha: list of p numeric matrices of dimension n \times (p - 1); the (i, j) entry of the k-th matrix is the variational approximation to the posterior inclusion probability of the j-th variable in a weighted regression with variable k fixed as the response, where the weights are taken with respect to observation i mu: list of p numeric matrices of dimension n \times (p - 1); the (i, j) entry of the k-th matrix is the variational approximation to the posterior slab mean for the j-th variable in a weighted regression with variable k fixed as the response, where the weights are taken with respect to observation i ssq_var: list of p numeric matrices of dimension n \times (p - 1); the (i, j) entry of the k-th matrix is the variational approximation to the posterior slab variance for the j-th variable in a weighted regression with variable k fixed as the response, where the weights are taken with respect to observation i
hyperparameters	list of p lists; the j-th list has the following values for variable j fixed as the response: grid: matrix of candidate hyperparameter values, corresponding ELBO, and iterations to converge final: the final hyperparameters chosen by grid search and the ELBO and iterations to converge for these hyperparameters
model_details	list with the following values: elapsed: amount of time to fit the model n: number of observations p: number of variables ELBO: ELBO summed across all observations and variables. If hp_method is "model_average" or "hybrid", this ELBO is averaged across the hyperparameter grid using the model averaging weights for each variable num_unique: number of unique graphs grid_size: number of points in the hyperparameter grid args: list containing all passed arguments of length 1
weights	list with the following values: weights: n\times n numeric matrix. The (i, j) entry is the similarity weight of the i-th observation with respect to the j-th observation using the j-th observation's bandwidth bandwidths: numeric vector of length n. The i-th entry is the bandwidth for the i-th observation

References

(1) Sutanoy Dasgupta, Peng Zhao, Jacob Helwig, Prasenjit Ghosh, Debdeep Pati, and Bani Mallick. An Approximate Bayesian Approach to Covariate-dependent Graphical Modeling. arXiv preprint, 1–64, 2023.

(2) Sutanoy Dasgupta, Debdeep Pati, and Anuj Srivastava. A Two-Step Geometric Framework For Density Modeling. Statistica Sinica, 30(4):2155–2177, 2020.

Examples

## Not run: 
library(ggplot2)

# get the data
set.seed(12)
data <- generateData()
X <- data$X
Z <- data$Z
interval <- data$interval
prec <- data$true_precision

# get overall and within interval sample sizes
n <- nrow(X)
n1 <- sum(interval == 1)
n2 <- sum(interval == 2)
n3 <- sum(interval == 3)

# visualize the distribution of the extraneous covariate
ggplot(data.frame(Z = Z, interval = as.factor(interval))) +
  geom_histogram(aes(Z, fill = interval), color = "black", bins = n %/% 5)

# visualize the true precision matrices in each of the intervals

# interval 1
matViz(prec[[1]], incl_val = TRUE) +
  ggtitle(paste0("True precision matrix, interval 1, observations 1,...,", n1))

# interval 2 (varies continuously with Z)
cat("\nInterval 2, observations ", n1 + 1, ",...,", n1 + n2, sep = "")
int2_mats <- prec[interval == 2]
int2_inds <- c(5, n2 %/% 2, n2 - 5)
lapply(int2_inds, function(j) matViz(int2_mats[[j]], incl_val = TRUE) +
         ggtitle(paste("True precision matrix, interval 2, observation", j + n1)))

# interval 3
matViz(prec[[length(prec)]], incl_val = TRUE) +
  ggtitle(paste0("True precision matrix, interval 3, observations ",
                 n1 + n2 + 1, ",...,", n1 + n2 + n3))

# fit the model and visualize the estimated graphs
(out <- covdepGE(X, Z))
plot(out)

# visualize the posterior inclusion probabilities for variables (1, 3) and (1, 2)
inclusionCurve(out, 1, 2)
inclusionCurve(out, 1, 3)

## End(Not run)

JacobHelwig/covdepGE documentation built on April 11, 2024, 7:22 a.m.