Title: Bayesian Predictive Stacking for Scalable Geospatial Transfer Learning
Version: 0.0-4
Maintainer: Luca Presicce <l.presicce@campus.unimib.it>
Author: Luca Presicce ORCID iD [aut, cre], Sudipto Banerjee [aut]
Description: Provides functions for Bayesian Predictive Stacking within the Bayesian transfer learning framework for geospatial artificial systems, as introduced in "Bayesian Transfer Learning for Artificially Intelligent Geospatial Systems: A Predictive Stacking Approach" (Presicce and Banerjee, 2024) <doi:10.48550/arXiv.2410.09504>. This methodology enables efficient Bayesian geostatistical modeling, utilizing predictive stacking to improve inference across spatial datasets. The core functions leverage 'C++' for high-performance computation, making the framework well-suited for large-scale spatial data analysis in parallel and distributed computing environments. Designed for scalability, it allows seamless application in computationally demanding scenarios.
Depends: R (≥ 1.8.0)
Imports: Rcpp, CVXR, mniw
LinkingTo: Rcpp, RcppArmadillo
Suggests: knitr, rmarkdown, mvnfast, foreach, parallel, doParallel, tictoc, MBA, RColorBrewer, classInt, sp, fields, testthat (≥ 3.0.0)
Config/testthat/edition: 3
License: GPL (≥ 3)
Encoding: UTF-8
RoxygenNote: 7.3.1
VignetteBuilder: knitr
NeedsCompilation: yes
Packaged: 2024-10-24 16:20:38 UTC; presi
Repository: CRAN
Date/Publication: 2024-10-25 09:20:01 UTC

spBPS: Bayesian Predictive Stacking for Scalable Geospatial Transfer Learning

Description

Provides functions for Bayesian Predictive Stacking within the Bayesian transfer learning framework for geospatial artificial systems, as introduced in "Bayesian Transfer Learning for Artificially Intelligent Geospatial Systems: A Predictive Stacking Approach" (Presicce and Banerjee, 2024) doi:10.48550/arXiv.2410.09504. This methodology enables efficient Bayesian geostatistical modeling, utilizing predictive stacking to improve inference across spatial datasets. The core functions leverage 'C++' for high-performance computation, making the framework well-suited for large-scale spatial data analysis in parallel and distributed computing environments. Designed for scalability, it allows seamless application in computationally demanding scenarios.

Author(s)

Maintainer: Luca Presicce l.presicce@campus.unimib.it (ORCID)

Authors:

Combine subset models wiht Pseudo-BMA

Description

Combine subset models wiht Pseudo-BMA

Usage

BPS_PseudoBMA(fit_list)

Arguments

fit_list

list K fitted model outputs composed by two elements each: first named epd, second named W

Value

matrix posterior predictive density evaluations (each columns represent a different model)

Examples


## Generate subsets of data
n <- 100
p <- 3
X <- matrix(rnorm(n*p), nrow = n, ncol = p)
Y <- matrix(rnorm(n), nrow = n, ncol = 1)
crd <- matrix(runif(n*2), nrow = n, ncol = 2)
data_part <- subset_data(data = list(Y = Y, X = X, crd = crd), K = 10)

## Select competitive set of values for hyperparameters
delta_seq <- c(0.1, 0.2, 0.3)
phi_seq <- c(3, 4, 5)

## Perform Bayesian Predictive Stacking within subsets
fit_list <- vector(length = 10, mode = "list")
for (i in 1:10) {
    Yi <- data_part$Y_list[[i]]
    Xi <- data_part$X_list[[i]]
    crd_i <- data_part$crd_list[[i]]
    p <- ncol(Xi)
    bps <- spBPS::BPS_weights(data = list(Y = Yi, X = Xi),
                               priors = list(mu_b = matrix(rep(0, p)),
                                             V_b = diag(10, p),
                                             a = 2,
                                             b = 2), coords = crd_i,
                                             hyperpar = list(delta = delta_seq,
                                                             phi = phi_seq),
                                             K = 5)
     w_hat <- bps$W
     epd <- bps$epd
     fit_list[[i]] <- list(epd, w_hat) }

## Combination weights between partitions using Pseudo Bayesian Model Averaging
comb_bps <- BPS_PseudoBMA(fit_list = fit_list)

Combine subset models wiht BPS

Description

Combine subset models wiht BPS

Usage

BPS_combine(fit_list, K, rp)

Arguments

fit_list

list K fitted model outputs composed by two elements each: first named epd, second named W

K

integer number of folds

rp

double percentage of observations to take into account for optimization (default=1)

Value

matrix posterior predictive density evaluations (each columns represent a different model)

Examples


## Generate subsets of data
n <- 100
p <- 3
X <- matrix(rnorm(n*p), nrow = n, ncol = p)
Y <- matrix(rnorm(n), nrow = n, ncol = 1)
crd <- matrix(runif(n*2), nrow = n, ncol = 2)
data_part <- subset_data(data = list(Y = Y, X = X, crd = crd), K = 10)

## Select competitive set of values for hyperparameters
delta_seq <- c(0.1, 0.2, 0.3)
phi_seq <- c(3, 4, 5)

## Perform Bayesian Predictive Stacking within subsets
fit_list <- vector(length = 10, mode = "list")
for (i in 1:10) {
    Yi <- data_part$Y_list[[i]]
    Xi <- data_part$X_list[[i]]
    crd_i <- data_part$crd_list[[i]]
    p <- ncol(Xi)
    bps <- spBPS::BPS_weights(data = list(Y = Yi, X = Xi),
                               priors = list(mu_b = matrix(rep(0, p)),
                                             V_b = diag(10, p),
                                             a = 2,
                                             b = 2), coords = crd_i,
                                             hyperpar = list(delta = delta_seq,
                                                             phi = phi_seq),
                                             K = 5)
     w_hat <- bps$W
     epd <- bps$epd
     fit_list[[i]] <- list(epd, w_hat) }

## Combination weights between partitions using Bayesian Predictive Stacking
comb_bps <- BPS_combine(fit_list = fit_list, K = 10, rp = 1)

Perform the BPS sampling from posterior and posterior predictive given a set of stacking weights

Description

Perform the BPS sampling from posterior and posterior predictive given a set of stacking weights

Usage

BPS_post(data, X_u, priors, coords, crd_u, hyperpar, W, R)

Arguments

data

list two elements: first named Y, second named X

X_u

matrix unobserved instances covariate matrix

priors

list priors: named \mu_b,V_b,a,b

coords

matrix sample coordinates for X and Y

crd_u

matrix unboserved instances coordinates

hyperpar

list two elemets: first named \delta, second named \phi

W

matrix set of stacking weights

R

integer number of desired samples

Value

list BPS posterior predictive samples

Examples


## Generate subsets of data
n <- 100
p <- 3
X <- matrix(rnorm(n*p), nrow = n, ncol = p)
Y <- matrix(rnorm(n), nrow = n, ncol = 1)
crd <- matrix(runif(n*2), nrow = n, ncol = 2)
data_part <- subset_data(data = list(Y = Y, X = X, crd = crd), K = 10)

## Select competetive set of values for hyperparameters
delta_seq <- c(0.1, 0.2, 0.3)
phi_seq <- c(3, 4, 5)
## Fit local models
fit_list <- vector(length = 10, mode = "list")
for (i in 1:10) {
  Yi <- data_part$Y_list[[i]]
  Xi <- data_part$X_list[[i]]
  crd_i <- data_part$crd_list[[i]]
  p <- ncol(Xi)
  bps <- spBPS::BPS_weights(data = list(Y = Yi, X = Xi),
                           priors = list(mu_b = matrix(rep(0, p)),
                                         V_b = diag(10, p),
                                         a = 2,
                                         b = 2), coords = crd_i,
                                         hyperpar = list(delta = delta_seq,
                                                         phi = phi_seq),
                                                         K = 5)
  w_hat <- bps$W
  epd <- bps$epd
  fit_list[[i]] <- list(epd, w_hat) }

## Model combination weights between partitions using Bayesian Predictive Stacking
comb_bps <- BPS_combine(fit_list = fit_list, K = 10, rp = 1)
Wbps <- comb_bps$W
W_list <- comb_bps$W_list

## Generate prediction points
m <- 100
X_new <- matrix(rnorm(m*p), nrow = m, ncol = p)
crd_new <- matrix(runif(m*2), nrow = m, ncol = 2)

## Perform posterior and posterior predictive sampling
R <- 250
subset_ind <- sample(1:10, R, TRUE, Wbps)
postsmp_and_pred <- vector(length = R, mode = "list")
for (r in 1:R) {
 ind_s <- subset_ind[r]
 Ys <- matrix(data_part$Y_list[[ind_s]])
 Xs <- data_part$X_list[[ind_s]]
 crds <- data_part$crd_list[[ind_s]]
 Ws <- W_list[[ind_s]]
 result <- spBPS::BPS_post(data = list(Y = Ys, X = Xs), coords = crds,
                           X_u = X_new, crd_u = crd_new,
                           priors = list(mu_b = matrix(rep(0, p)),
                                         V_b = diag(10, p),
                                         a = 2,
                                         b = 2),
                                         hyperpar = list(delta = delta_seq,
                                                         phi = phi_seq),
                                                         W = Ws, R = 1)
 postsmp_and_pred[[r]] <- result}

Perform the BPS sampling from posterior and posterior predictive given a set of stacking weights

Description

Perform the BPS sampling from posterior and posterior predictive given a set of stacking weights

Usage

BPS_post_MvT(data, X_u, priors, coords, crd_u, hyperpar, W, R)

Arguments

data

list two elements: first named Y, second named X

X_u

matrix unobserved instances covariate matrix

priors

list priors: named \mu_B,V_r,\Psi,\nu

coords

matrix sample coordinates for X and Y

crd_u

matrix unboserved instances coordinates

hyperpar

list two elemets: first named \alpha, second named \phi

W

matrix set of stacking weights

R

integer number of desired samples

Value

list BPS posterior predictive samples

Examples


## Generate subsets of data
n <- 100
p <- 3
q <- 2
X <- matrix(rnorm(n*p), nrow = n, ncol = p)
Y <- matrix(rnorm(n*q), nrow = n, ncol = q)
crd <- matrix(runif(n*2), nrow = n, ncol = 2)
data_part <- subset_data(data = list(Y = Y, X = X, crd = crd), K = 10)

## Select competitive set of values for hyperparameters
alfa_seq <- c(0.7, 0.8, 0.9)
phi_seq <- c(3, 4, 5)

## Fit local models
fit_list <- vector(length = 10, mode = "list")
for (i in 1:10) {
    Yi <- data_part$Y_list[[i]]
    Xi <- data_part$X_list[[i]]
    crd_i <- data_part$crd_list[[i]]
    bps <- spBPS::BPS_weights_MvT(data = list(Y = Yi, X = Xi),
                              priors = list(mu_B = matrix(0, nrow = p, ncol = q),
                                            V_r = diag(10, p),
                                            Psi = diag(1, q),
                                            nu = 3), coords = crd_i,
                                            hyperpar = list(alpha = alfa_seq,
                                                            phi = phi_seq),
                                            K = 5)
     w_hat <- bps$W
     epd <- bps$epd
     fit_list[[i]] <- list(epd, w_hat) }

## Model combination weights between partitions using Bayesian Predictive Stacking
comb_bps <- BPS_combine(fit_list = fit_list, K = 10, rp = 1)
Wbps <- comb_bps$W
W_list <- comb_bps$W_list

## Generate prediction points
m <- 100
X_new <- matrix(rnorm(m*p), nrow = m, ncol = p)
crd_new <- matrix(runif(m*2), nrow = m, ncol = 2)

## Perform posterior and posterior predictive sampling
R <- 250
subset_ind <- sample(1:10, R, TRUE, Wbps)
postsmp_and_pred <- vector(length = R, mode = "list")
for (r in 1:R) {
  ind_s <- subset_ind[r]
  Ys <- data_part$Y_list[[ind_s]]
  Xs <- data_part$X_list[[ind_s]]
  crds <- data_part$crd_list[[ind_s]]
  Ws <- W_list[[ind_s]]
  result <- spBPS::BPS_post_MvT(data = list(Y = Ys, X = Xs), coords = crds,
                                X_u = X_new, crd_u = crd_new,
                                priors = list(mu_B = matrix(0, nrow = p, ncol = q),
                                              V_r = diag(10, p),
                                              Psi = diag(1, q),
                                              nu = 3),
                                              hyperpar = list(alpha = alfa_seq,
                                                              phi = phi_seq),
                                              W = Ws, R = 1)

  postsmp_and_pred[[r]] <- result}

Compute the BPS posterior samples given a set of stacking weights

Description

Compute the BPS posterior samples given a set of stacking weights

Usage

BPS_postdraws(data, priors, coords, hyperpar, W, R)

Arguments

data

list two elements: first named Y, second named X

priors

list priors: named \mu_b,V_b,a,b

coords

matrix sample coordinates for X and Y

hyperpar

list two elemets: first named \delta, second named \phi

W

matrix set of stacking weights

R

integer number of desired samples

Value

matrix BPS posterior samples

Compute the BPS posterior samples given a set of stacking weights

Description

Compute the BPS posterior samples given a set of stacking weights

Usage

BPS_postdraws_MvT(data, priors, coords, hyperpar, W, R, par)

Arguments

data

list two elements: first named Y, second named X

priors

list priors: named \mu_B,V_r,\Psi,\nu

coords

matrix sample coordinates for X and Y

hyperpar

list two elemets: first named \alpha, second named \phi

W

matrix set of stacking weights

R

integer number of desired samples

par

if TRUE only \beta and \Sigma are sampled (\omega is omitted)

Value

matrix BPS posterior samples

Compute the BPS spatial prediction given a set of stacking weights

Description

Compute the BPS spatial prediction given a set of stacking weights

Usage

BPS_pred(data, X_u, priors, coords, crd_u, hyperpar, W, R)

Arguments

data

list two elements: first named Y, second named X

X_u

matrix unobserved instances covariate matrix

priors

list priors: named \mu_b,V_b,a,b

coords

matrix sample coordinates for X and Y

crd_u

matrix unboserved instances coordinates

hyperpar

list two elemets: first named \delta, second named \phi

W

matrix set of stacking weights

R

integer number of desired samples

Value

list BPS posterior predictive samples

Examples


## Generate subsets of data
n <- 100
p <- 3
X <- matrix(rnorm(n*p), nrow = n, ncol = p)
Y <- matrix(rnorm(n), nrow = n, ncol = 1)
crd <- matrix(runif(n*2), nrow = n, ncol = 2)
data_part <- subset_data(data = list(Y = Y, X = X, crd = crd), K = 10)

## Select competetive set of values for hyperparameters
delta_seq <- c(0.1, 0.2, 0.3)
phi_seq <- c(3, 4, 5)

## Fit local models
fit_list <- vector(length = 10, mode = "list")
for (i in 1:10) {
    Yi <- data_part$Y_list[[i]]
    Xi <- data_part$X_list[[i]]
    crd_i <- data_part$crd_list[[i]]
    p <- ncol(Xi)
    bps <- spBPS::BPS_weights(data = list(Y = Yi, X = Xi),
                               priors = list(mu_b = matrix(rep(0, p)),
                                             V_b = diag(10, p),
                                             a = 2,
                                             b = 2), coords = crd_i,
                                             hyperpar = list(delta = delta_seq,
                                                             phi = phi_seq),
                                             K = 5)
     w_hat <- bps$W
     epd <- bps$epd
     fit_list[[i]] <- list(epd, w_hat) }

## Model combination weights between partitions using Bayesian Predictive Stacking
comb_bps <- BPS_combine(fit_list = fit_list, K = 10, rp = 1)
Wbps <- comb_bps$W
W_list <- comb_bps$W_list

## Generate prediction points
m <- 50
X_new <- matrix(rnorm(m*p), nrow = m, ncol = p)
crd_new <- matrix(runif(m*2), nrow = m, ncol = 2)

## Perform posterior predictive sampling
R <- 250
subset_ind <- sample(1:10, R, TRUE, Wbps)
predictions <- vector(length = R, mode = "list")
for (r in 1:R) {
  ind_s <- subset_ind[r]
  Ys <- matrix(data_part$Y_list[[ind_s]])
  Xs <- data_part$X_list[[ind_s]]
  crds <- data_part$crd_list[[ind_s]]
  Ws <- W_list[[ind_s]]
  result <- spBPS::BPS_pred(data = list(Y = Ys, X = Xs), coords = crds,
                            X_u = X_new, crd_u = crd_new,
                            priors = list(mu_b = matrix(rep(0, p)),
                                          V_b = diag(10, p),
                                          a = 2,
                                          b = 2),
                                          hyperpar = list(delta = delta_seq,
                                                          phi = phi_seq),
                                          W = Ws, R = 1)

  predictions[[r]] <- result}

Compute the BPS spatial prediction given a set of stacking weights

Description

Compute the BPS spatial prediction given a set of stacking weights

Usage

BPS_pred_MvT(data, X_u, priors, coords, crd_u, hyperpar, W, R)

Arguments

data

list two elements: first named Y, second named X

X_u

matrix unobserved instances covariate matrix

priors

list priors: named \mu_B,V_r,\Psi,\nu

coords

matrix sample coordinates for X and Y

crd_u

matrix unboserved instances coordinates

hyperpar

list two elemets: first named \alpha, second named \phi

W

matrix set of stacking weights

R

integer number of desired samples

Value

list BPS posterior predictive samples

Examples


## Generate subsets of data
n <- 100
p <- 3
q <- 2
X <- matrix(rnorm(n*p), nrow = n, ncol = p)
Y <- matrix(rnorm(n*q), nrow = n, ncol = q)
crd <- matrix(runif(n*2), nrow = n, ncol = 2)
data_part <- subset_data(data = list(Y = Y, X = X, crd = crd), K = 10)

## Select competitive set of values for hyperparameters
alfa_seq <- c(0.7, 0.8, 0.9)
phi_seq <- c(3, 4, 5)

## Fit local models
fit_list <- vector(length = 10, mode = "list")
for (i in 1:10) {
    Yi <- data_part$Y_list[[i]]
    Xi <- data_part$X_list[[i]]
    crd_i <- data_part$crd_list[[i]]
    bps <- spBPS::BPS_weights_MvT(data = list(Y = Yi, X = Xi),
                              priors = list(mu_B = matrix(0, nrow = p, ncol = q),
                                            V_r = diag(10, p),
                                            Psi = diag(1, q),
                                            nu = 3), coords = crd_i,
                                            hyperpar = list(alpha = alfa_seq,
                                                            phi = phi_seq),
                                            K = 5)
     w_hat <- bps$W
     epd <- bps$epd
     fit_list[[i]] <- list(epd, w_hat) }

## Model combination weights between partitions using Bayesian Predictive Stacking
comb_bps <- BPS_combine(fit_list = fit_list, K = 10, rp = 1)
Wbps <- comb_bps$W
W_list <- comb_bps$W_list

## Generate prediction points
m <- 100
X_new <- matrix(rnorm(m*p), nrow = m, ncol = p)
crd_new <- matrix(runif(m*2), nrow = m, ncol = 2)

## Perform posterior predictive sampling
R <- 250
subset_ind <- sample(1:10, R, TRUE, Wbps)
predictions <- vector(length = R, mode = "list")
for (r in 1:R) {
  ind_s <- subset_ind[r]
  Ys <- data_part$Y_list[[ind_s]]
  Xs <- data_part$X_list[[ind_s]]
  crds <- data_part$crd_list[[ind_s]]
  Ws <- W_list[[ind_s]]
  result <- spBPS::BPS_pred_MvT(data = list(Y = Ys, X = Xs), coords = crds,
                                X_u = X_new, crd_u = crd_new,
                                priors = list(mu_B = matrix(0, nrow = p, ncol = q),
                                              V_r = diag(10, p),
                                              Psi = diag(1, q),
                                              nu = 3),
                                              hyperpar = list(alpha = alfa_seq,
                                                              phi = phi_seq),
                                              W = Ws, R = 1)

  predictions[[r]] <- result}

Compute the BPS weights by convex optimization

Description

Compute the BPS weights by convex optimization

Usage

BPS_weights(data, priors, coords, hyperpar, K)

Arguments

data

list two elements: first named Y, second named X

priors

list priors: named \mu_b,V_b,a,b

coords

matrix sample coordinates for X and Y

hyperpar

list two elemets: first named \delta, second named \phi

K

integer number of folds

Value

matrix posterior predictive density evaluations (each columns represent a different model)

Examples

## Generate subsets of data
n <- 100
p <- 3
X <- matrix(rnorm(n*p), nrow = n, ncol = p)
Y <- matrix(rnorm(n), nrow = n)
crd <- matrix(runif(n*2), nrow = n, ncol = 2)

## Select competitive set of values for hyperparameters
delta_seq <- c(0.1, 0.2, 0.3)
phi_seq <- c(3, 4, 5)

## Perform Bayesian Predictive Stacking within subsets
bps <- spBPS::BPS_weights(data = list(Y = Y, X = X),
                               priors = list(mu_b = matrix(rep(0, p)),
                                             V_b = diag(10, p),
                                             a = 2,
                                             b = 2), coords = crd,
                                             hyperpar = list(delta = delta_seq,
                                                             phi = phi_seq),
                                             K = 5)

Compute the BPS weights by convex optimization

Description

Compute the BPS weights by convex optimization

Usage

BPS_weights_MvT(data, priors, coords, hyperpar, K)

Arguments

data

list two elements: first named Y, second named X

priors

list priors: named \mu_B,V_r,\Psi,\nu

coords

matrix sample coordinates for X and Y

hyperpar

list two elemets: first named \alpha, second named \phi

K

integer number of folds

Value

matrix posterior predictive density evaluations (each columns represent a different model)

Examples


## Generate subsets of data
n <- 100
p <- 3
q <- 2
X <- matrix(rnorm(n*p), nrow = n, ncol = p)
Y <- matrix(rnorm(n*q), nrow = n, ncol = q)
crd <- matrix(runif(n*2), nrow = n, ncol = 2)

## Select competitive set of values for hyperparameters
alfa_seq <- c(0.7, 0.8, 0.9)
phi_seq <- c(3, 4, 5)

## Perform Bayesian Predictive Stacking within subsets
bps <- spBPS::BPS_weights_MvT(data = list(Y = Y, X = X),
                              priors = list(mu_B = matrix(0, nrow = p, ncol = q),
                                            V_r = diag(10, p),
                                            Psi = diag(1, q),
                                            nu = 3), coords = crd,
                                            hyperpar = list(alpha = alfa_seq,
                                                            phi = phi_seq),
                                            K = 5)

Compute the BPS weights by convex optimization

Description

Compute the BPS weights by convex optimization

Usage

CVXR_opt(scores)

Arguments

scores

matrix N \times K of expected predictive density evaluations for the K models considered

Value

conv_opt function to perform convex optimiazion with CVXR R package

Compute the Euclidean distance matrix

Description

Compute the Euclidean distance matrix

Usage

arma_dist(X)

Arguments

X

matrix (tipically of N coordindates on \mathbb{R}^2 )

Value

matrix distance matrix of the elements of X

Examples

## Compute the Distance matrix of dimension (n x n)
n <- 100
p <- 2
X <- matrix(runif(n*p), nrow = n, ncol = p)
distance.matrix <- arma_dist(X)

Gibbs sampler for Conjugate Bayesian Multivariate Linear Models

Description

Gibbs sampler for Conjugate Bayesian Multivariate Linear Models

Usage

bayesMvLMconjugate(Y, X, mu_B, V_B, nu, Psi, n_iter = 1000, burn_in = 500)

Arguments

Y

matrix n \times q of response variables

X

matrix n \times p of predictors

mu_B

matrix p \times q prior mean for \beta

V_B

matrix p \times p prior row covariance for \beta

nu

double prior parameter for \Sigma

Psi

matrix prior parameter for \Sigma

n_iter

integer iteration number for Gibbs sampler

burn_in

integer number of burn-in iteration

Value

B_samples array of posterior sample for \beta

Sigma_samples array of posterior samples for \Sigma

Examples

## Generate data
n <- 100
p <- 3
q <- 2
Y <- matrix(rnorm(n*q), nrow = n, ncol = q)
X <- matrix(rnorm(n*p), nrow = n, ncol = p)

## Prior parameters
mu_B <- matrix(0, p, q)
V_B <- diag(10, p)
nu <- 3
Psi <- diag(q)

## Samples from posteriors
n_iter <- 1000
burn_in <- 500
set.seed(1234)
samples <- spBPS::bayesMvLMconjugate(Y, X, mu_B, V_B, nu, Psi, n_iter, burn_in)

Solver for Bayesian Predictive Stacking of Predictive densities convex optimization problem

Description

Solver for Bayesian Predictive Stacking of Predictive densities convex optimization problem

Usage

conv_opt(scores)

Arguments

scores

matrix N \times K of expected predictive density evaluations for the K models considered

Value

W matrix of Bayesian Predictive Stacking weights for the K models considered

Examples

## Generate (randomly) K predictive scores for n observations
n <- 50
K <- 5
scores <- matrix(runif(n*K), nrow = n, ncol = K)

## Find Bayesian Predictive Stacking weights
opt_weights <- conv_opt(scores)

Evaluate the density of a set of unobserved response with respect to the conditional posterior predictive

Description

Evaluate the density of a set of unobserved response with respect to the conditional posterior predictive

Usage

d_pred_cpp(data, X_u, Y_u, d_u, d_us, hyperpar, poster)

Arguments

data

list two elements: first named Y, second named X

X_u

matrix unobserved instances covariate matrix

Y_u

matrix unobserved instances response matrix

d_u

matrix unobserved instances distance matrix

d_us

matrix cross-distance between unobserved and observed instances matrix

hyperpar

list two elemets: first named \delta, second named \phi

poster

list output from fit_cpp function

Value

vector posterior predictive density evaluations

Evaluate the density of a set of unobserved response with respect to the conditional posterior predictive

Description

Evaluate the density of a set of unobserved response with respect to the conditional posterior predictive

Usage

d_pred_cpp_MvT(data, X_u, Y_u, d_u, d_us, hyperpar, poster)

Arguments

data

list two elements: first named Y, second named X

X_u

matrix unobserved instances covariate matrix

Y_u

matrix unobserved instances response matrix

d_u

matrix unobserved instances distance matrix

d_us

matrix cross-distance between unobserved and observed instances matrix

hyperpar

list two elemets: first named \alpha, second named \phi

poster

list output from fit_cpp function

Value

double posterior predictive density evaluation

Compute the KCV of the density evaluations for fixed values of the hyperparameters

Description

Compute the KCV of the density evaluations for fixed values of the hyperparameters

Usage

dens_kcv(data, priors, coords, hyperpar, K)

Arguments

data

list two elements: first named Y, second named X

priors

list priors: named \mu_b,V_b,a,b

coords

matrix sample coordinates for X and Y

hyperpar

list two elemets: first named \delta, second named \phi

K

integer number of folds

Value

vector posterior predictive density evaluations

Compute the KCV of the density evaluations for fixed values of the hyperparameters

Description

Compute the KCV of the density evaluations for fixed values of the hyperparameters

Usage

dens_kcv_MvT(data, priors, coords, hyperpar, K)

Arguments

data

list two elements: first named Y, second named X

priors

list priors: named \mu_B,V_r,\Psi,\nu

coords

matrix sample coordinates for X and Y

hyperpar

list two elemets: first named \alpha, second named \phi

K

integer number of folds

Value

vector posterior predictive density evaluations

Compute the LOOCV of the density evaluations for fixed values of the hyperparameters

Description

Compute the LOOCV of the density evaluations for fixed values of the hyperparameters

Usage

dens_loocv(data, priors, coords, hyperpar)

Arguments

data

list two elements: first named Y, second named X

priors

list priors: named \mu_b,V_b,a,b

coords

matrix sample coordinates for X and Y

hyperpar

list two elemets: first named \delta, second named \phi

Value

vector posterior predictive density evaluations

Compute the LOOCV of the density evaluations for fixed values of the hyperparameters

Description

Compute the LOOCV of the density evaluations for fixed values of the hyperparameters

Usage

dens_loocv_MvT(data, priors, coords, hyperpar)

Arguments

data

list two elements: first named Y, second named X

priors

list priors: named \mu_B,V_r,\Psi,\nu

coords

matrix sample coordinates for X and Y

hyperpar

list two elemets: first named \alpha, second named \phi

Value

vector posterior predictive density evaluations

Build a grid from two vector (i.e. equivalent to expand.grid() in R)

Description

Build a grid from two vector (i.e. equivalent to expand.grid() in R)

Usage

expand_grid_cpp(x, y)

Arguments

x

vector first vector of numeric elements

y

vector second vector of numeric elements

Value

matrix expanded grid of combinations

Examples

## Create a matrix from all combination of vectors
x <- seq(0, 10, length.out = 100)
y <- seq(-1, 1, length.out = 20)
grid <- expand_grid_cpp(x = x, y = y)

Compute the parameters for the posteriors distribution of \beta and \Sigma (i.e. updated parameters)

Description

Compute the parameters for the posteriors distribution of \beta and \Sigma (i.e. updated parameters)

Usage

fit_cpp(data, priors, coords, hyperpar)

Arguments

data

list two elements: first named Y, second named X

priors

list priors: named \mu_b,V_b,a,b

coords

matrix sample coordinates for X and Y

hyperpar

list two elemets: first named \delta, second named \phi

Value

list posterior update parameters

Compute the parameters for the posteriors distribution of \beta and \Sigma (i.e. updated parameters)

Description

Compute the parameters for the posteriors distribution of \beta and \Sigma (i.e. updated parameters)

Usage

fit_cpp_MvT(data, priors, coords, hyperpar)

Arguments

data

list two elements: first named Y, second named X

priors

list priors: named \mu_B,V_r,\Psi,\nu

coords

matrix sample coordinates for X and Y

hyperpar

list two elemets: first named \alpha, second named \phi

Value

list posterior update parameters

Function to subset data for meta-analysis

Description

Function to subset data for meta-analysis

Usage

forceSymmetry_cpp(mat)

Arguments

mat

matrix not-symmetric matrix

Value

matrix symmetric matrix (lower triangular of mat is used)

Examples

## Force matrix to be symmetric (avoiding numerical problems)
n <- 4
X <- matrix(runif(n*n), nrow = n, ncol = n)
X <- forceSymmetry_cpp(mat = X)

Return the CV predictive density evaluations for all the model combinations

Description

Return the CV predictive density evaluations for all the model combinations

Usage

models_dens(data, priors, coords, hyperpar, useKCV, K)

Arguments

data

list two elements: first named Y, second named X

priors

list priors: named \mu_b,V_b,a,b

coords

matrix sample coordinates for X and Y

hyperpar

list two elemets: first named \delta, second named \phi

useKCV

if TRUE K-fold cross validation is used instead of LOOCV (no default)

K

integer number of folds

Value

matrix posterior predictive density evaluations (each columns represent a different model)

Return the CV predictive density evaluations for all the model combinations

Description

Return the CV predictive density evaluations for all the model combinations

Usage

models_dens_MvT(data, priors, coords, hyperpar, useKCV, K)

Arguments

data

list two elements: first named Y, second named X

priors

list priors: named \mu_B,V_r,\Psi,\nu

coords

matrix sample coordinates for X and Y

hyperpar

list two elemets: first named \alpha, second named \phi

useKCV

if TRUE K-fold cross validation is used instead of LOOCV (no default)

K

integer number of folds

Value

matrix posterior predictive density evaluations (each columns represent a different model)

Sample R draws from the posterior distributions

Description

Sample R draws from the posterior distributions

Usage

post_draws(poster, R, par, p)

Arguments

poster

list output from fit_cpp function

R

integer number of posterior samples

par

if TRUE only \beta and \sigma^2 are sampled (\omega is omitted)

p

integer if par = TRUE, it specifies the column number of X

Value

list posterior samples

Sample R draws from the posterior distributions

Description

Sample R draws from the posterior distributions

Usage

post_draws_MvT(poster, R, par, p)

Arguments

poster

list output from fit_cpp function

R

integer number of posterior samples

par

if TRUE only \beta and \Sigma are sampled (\omega is omitted)

p

integer if par = TRUE, it specifies the column number of X

Value

list posterior samples

Predictive sampler for Conjugate Bayesian Multivariate Linear Models

Description

Predictive sampler for Conjugate Bayesian Multivariate Linear Models

Usage

pred_bayesMvLMconjugate(X_new, B_samples, Sigma_samples)

Arguments

X_new

matrix n_new \times p of predictors for new data points

B_samples

array of posterior sample for \beta

Sigma_samples

array of posterior samples for \Sigma

Value

Y_pred matrix of posterior mean for response matrix Y predictions

Y_pred_samples array of posterior predictive sample for response matrix Y

Examples

## Generate data
n <- 100
p <- 3
q <- 2
Y <- matrix(rnorm(n*q), nrow = n, ncol = q)
X <- matrix(rnorm(n*p), nrow = n, ncol = p)

## Prior parameters
mu_B <- matrix(0, p, q)
V_B <- diag(10, p)
nu <- 3
Psi <- diag(q)

## Samples from posteriors
n_iter <- 1000
burn_in <- 500
set.seed(1234)
samples <- spBPS::bayesMvLMconjugate(Y, X, mu_B, V_B, nu, Psi, n_iter, burn_in)

## Extract posterior samples
B_samples <- samples$B_samples
Sigma_samples <- samples$Sigma_samples

## Samples from predictive posterior (based posterior samples)
m <- 50
X_new <- matrix(rnorm(m*p), nrow = m, ncol = p)
pred <- spBPS::pred_bayesMvLMconjugate(X_new, B_samples, Sigma_samples)

Draw from the conditional posterior predictive for a set of unobserved covariates

Description

Draw from the conditional posterior predictive for a set of unobserved covariates

Usage

r_pred_cond(data, X_u, d_u, d_us, hyperpar, poster, post)

Arguments

data

list two elements: first named Y, second named X

X_u

matrix unobserved instances covariate matrix

d_u

matrix unobserved instances distance matrix

d_us

matrix cross-distance between unobserved and observed instances matrix

hyperpar

list two elemets: first named \delta, second named \phi

poster

list output from fit_cpp function

post

list output from post_draws function

Value

list posterior predictive samples

Draw from the conditional posterior predictive for a set of unobserved covariates

Description

Draw from the conditional posterior predictive for a set of unobserved covariates

Usage

r_pred_cond_MvT(data, X_u, d_u, d_us, hyperpar, poster, post)

Arguments

data

list two elements: first named Y, second named X

X_u

matrix unobserved instances covariate matrix

d_u

matrix unobserved instances distance matrix

d_us

matrix cross-distance between unobserved and observed instances matrix

hyperpar

list two elemets: first named \alpha, second named \phi

poster

list output from fit_cpp_MvT function

post

list output from post_draws_MvT function

Value

list posterior predictive samples

Draw from the joint posterior predictive for a set of unobserved covariates

Description

Draw from the joint posterior predictive for a set of unobserved covariates

Usage

r_pred_joint(data, X_u, d_u, d_us, hyperpar, poster, R)

Arguments

data

list two elements: first named Y, second named X

X_u

matrix unobserved instances covariate matrix

d_u

matrix unobserved instances distance matrix

d_us

matrix cross-distance between unobserved and observed instances matrix

hyperpar

list two elemets: first named \delta, second named \phi

poster

list output from fit_cpp function

R

integer number of posterior predictive samples

Value

list posterior predictive samples

Draw from the joint posterior predictive for a set of unobserved covariates

Description

Draw from the joint posterior predictive for a set of unobserved covariates

Usage

r_pred_joint_MvT(data, X_u, d_u, d_us, hyperpar, poster, R)

Arguments

data

list two elements: first named Y, second named X

X_u

matrix unobserved instances covariate matrix

d_u

matrix unobserved instances distance matrix

d_us

matrix cross-distance between unobserved and observed instances matrix

hyperpar

list two elemets: first named \alpha, second named \phi

poster

list output from fit_cpp function

R

integer number of posterior predictive samples

Value

list posterior predictive samples

Draw from the marginals posterior predictive for a set of unobserved covariates

Description

Draw from the marginals posterior predictive for a set of unobserved covariates

Usage

r_pred_marg(data, X_u, d_u, d_us, hyperpar, poster, R)

Arguments

data

list two elements: first named Y, second named X

X_u

matrix unobserved instances covariate matrix

d_u

matrix unobserved instances distance matrix

d_us

matrix cross-distance between unobserved and observed instances matrix

hyperpar

list two elemets: first named \delta, second named \phi

poster

list output from fit_cpp function

R

integer number of posterior predictive samples

Value

list posterior predictive samples

Draw from the joint posterior predictive for a set of unobserved covariates

Description

Draw from the joint posterior predictive for a set of unobserved covariates

Usage

r_pred_marg_MvT(data, X_u, d_u, d_us, hyperpar, poster, R)

Arguments

data

list two elements: first named Y, second named X

X_u

matrix unobserved instances covariate matrix

d_u

matrix unobserved instances distance matrix

d_us

matrix cross-distance between unobserved and observed instances matrix

hyperpar

list two elemets: first named \alpha, second named \phi

poster

list output from fit_cpp function

R

integer number of posterior predictive samples

Value

list posterior predictive samples

Function to sample integers (index)

Description

Function to sample integers (index)

Usage

sample_index(size, length, p)

Arguments

size

integer dimension of the set to sample

length

integer number of elements to sample

p

vector sampling probabilities

Value

vector sample of integers

Perform prediction for BPS accelerated models - loop over prediction set

Description

Perform prediction for BPS accelerated models - loop over prediction set

Usage

spPredict_BPS(data, X_u, priors, coords, crd_u, hyperpar, W, R, J)

Arguments

data

list two elements: first named Y, second named X

X_u

matrix unobserved instances covariate matrix

priors

list priors: named \mu_b,V_b,a,b

coords

matrix sample coordinates for X and Y

crd_u

matrix unboserved instances coordinates

hyperpar

list two elemets: first named \delta, second named \phi

W

matrix set of stacking weights

R

integer number of desired samples

J

integer number of desired partition of prediction set

Value

list BPS posterior predictive samples

Function to subset data for meta-analysis

Description

Function to subset data for meta-analysis

Usage

subset_data(data, K)

Arguments

data

list three elements: first named Y, second named X, third named crd

K

integer number of desired subsets

Value

list subsets of data, and the set of indexes

Examples

## Create a list of K random subsets given a list with Y, X, and crd
n <- 100
p <- 3
q <- 2
X <- matrix(rnorm(n*p), nrow = n, ncol = p)
Y <- matrix(rnorm(n*q), nrow = n, ncol = q)
crd <- matrix(runif(n*2), nrow = n, ncol = 2)
subsets <- subset_data(data = list(Y = Y, X = X, crd = crd), K = 10)