Mixture-of-Experts Wishart Models for Covariance Data.

Mixture-of-Experts Wishart models

This R-package moewishart provides maximum likelihood estimation (MLE) and Bayesian estimation for the Wishart mixture model and the Wishart mixture-of-experts (MoE-Wishart) model. It implements four different inference algorithms for the two model:

• mixture model of Wishart distributions:
  • EM algorithm for finding the MLE;
  • Bayesian approach using Gibbs-within-MH sampling algorithm.
• Mixture-of-Expert model, in which the gating probabilities depend on covariates:
  • EM-MoE algorithm for finding the MLE;
  • Bayesian-MoE approach using a Gibbs-within-MH sampling algorithm.

Installation

Install the latest released version from CRAN:

install.packages("moewishart")

Install the latest development version from GitHub:

# library("devtools")
devtools::install_github("zhizuio/moewishart")

Example

Data simulation from a MoE-Wishart model:

• Sample size $n = 200$
• Dimension of the Wishart distribution $p = 2$
• Number of latent components $K = 3$
• $q=3$ gating covariates $\mathbf X = [x_{ij}] \in \mathbb R^{n\times q}$, $x_{ij}\sim\text{N}(0,1)$, $i=1,...,n$, $j=1,...,q$
• Fixed covariate effects $\boldsymbol\beta=[\boldsymbol\beta_1,...,\boldsymbol\beta_K] \in \mathbb R^{q\times K}$, with $\boldsymbol\beta_{K}=0$
• Probabilities of subpopulations $\boldsymbol\pi = [\pi_{ik}] \in \mathbb R^{n\times q}$, $\pi_{ik} = \exp(\mathbf X_i\boldsymbol\beta_k) / \sum_{l=1}^K\exp(\mathbf X_i\boldsymbol\beta_l)$, $k=1,...,K$
• Degrees of freedom $\boldsymbol\nu = (\nu_1,\nu_2,\nu_3) = (8, 12, 3)$
• Scale matrices of the Wishart distribution $\Sigma_1$, $\Sigma_2$, $\Sigma_3 \in \mathbb R^{p\times p}$
• Data $S_i \sim \pi_{i1}\text{Wishart}(\nu_1, \Sigma_1) + \pi_{i2}\text{Wishart}(\nu_2, \Sigma_2) + \pi_{i3}\text{Wishart}(\nu_3, \Sigma_3)$

1. Working model: Bayesian MoE-Wishart model

library(moewishart)

n <- 200 # number of subjects
p <- 2 # dimension of covariance matrix
set.seed(123) # fix coefficients of underlying MoE model
Xq <- 3
K <- 3
betas <- matrix(runif(Xq * K, -2, 2), nrow = Xq, ncol = K)
betas[, K] <- 0

# simulate data
dat <- simData(n, p,
  Xq = 3, K = 3, betas = betas,
  pis = c(0.35, 0.40, 0.25),
  nus = c(8, 12, 3)
)

# fit Bayesian MoE-Wishart model
set.seed(123)
fit <- moewishart(
  dat$S,
  X = cbind(1, dat$X), K = 3,
  mh_sigma = c(0.2, 0.1, 0.2), # RW-MH variances (length K)
  mh_beta = c(0.3, 0.3), # RW-MH variances (length K-1)
  niter = 3000, burnin = 1000
)

Posterior means for degrees of freedom (DoF) of Wishart distributions:

burnin <- 1000
nu_mcmc <- fit$nu[-c(1:burnin), ]
colMeans(nu_mcmc)
#> [1]  8.574911 14.397351  3.310689

True DoF:

dat$nu # true nu
#> [1]  8 12  3

Posterior means for scale matrices of Wishart distributions:

MoE_Sigma <- Reduce("+", fit$Sigma) / length(fit$Sigma)
MoE_Sigma
#> , , 1
#> 
#>           [,1]      [,2]
#> [1,] 0.5197160 0.2103881
#> [2,] 0.2103881 0.7470847
#> 
#> , , 2
#> 
#>           [,1]      [,2]
#> [1,] 1.7637949 0.5540576
#> [2,] 0.5540576 1.3244947
#> 
#> , , 3
#> 
#>            [,1]       [,2]
#> [1,]  4.1115070 -0.1267705
#> [2,] -0.1267705  3.0385263

Posterior means for gating coefficients:

beta_mcmc <- fit$Beta_samples[-c(1:burnin), , ]
apply(beta_mcmc, c(2, 3), mean)
#>            [,1]        [,2] [,3]
#> [1,] -0.3656861 -0.08024419    0
#> [2,] -0.9526224  2.24956385    0
#> [3,]  1.7609922  2.40287152    0
#> [4,] -0.4953755 -2.56072719    0

2. Working model: Bayesian Wishart mixture model

# fit Bayesian Wishart mixture model
set.seed(123)
fit2 <- mixturewishart(
  dat$S,
  K = 3,
  mh_sigma = c(0.2, 0.1, 0.2), # RW-MH variances
  niter = 3000, burnin = 1000
)

Posterior means for subpopulation probabilities:

colMeans(fit2$pi[-c(1:burnin), ])
#> [1] 0.2690425 0.5088864 0.2220712

Posterior means for DoF of Wishart distributions:

colMeans(fit2$nu[-c(1:burnin), ])
#> [1]  7.986113 12.153338  3.284252

3. Working model: MoE-Wishart model via EM algorithm

# fit MoE-Wishart model via EM alg.
set.seed(123)
fit3 <- moewishart(
  dat$S,
  X = cbind(1, dat$X), K = 3,
  method = "em",
  niter = 3000
)

EM estimates for DoF of Wishart distributions:

fit3$nu
#> [1]  7.515417 13.987158  3.274665

EM estimates for Wishart scale matrices:

fit3$Sigma
#> [[1]]
#>           [,1]      [,2]
#> [1,] 0.5591113 0.2324429
#> [2,] 0.2324429 0.8148737
#> 
#> [[2]]
#>           [,1]      [,2]
#> [1,] 1.7665723 0.5567668
#> [2,] 0.5567668 1.3336367
#> 
#> [[3]]
#>            [,1]       [,2]
#> [1,]  4.3139885 -0.1886288
#> [2,] -0.1886288  3.0983710

EM estimates for gating coefficients:

fit3$Beta
#>             comp1      comp2 comp3
#> [1,] -0.006270492  0.1302039     0
#> [2,] -0.798302303  2.0340525     0
#> [3,]  1.598530103  2.2399293     0
#> [4,] -0.510585695 -2.3465248     0

4. Working model: Wishart mixture model via EM algorithm

# fit Wishart mixture model via EM alg.
set.seed(123)
fit4 <- mixturewishart(
  dat$S,
  K = 3,
  method = "em",
  niter = 3000
)

EM estimates for DoF of Wishart distributions:

fit4$nu
#> [1]  2.995383  7.682819 11.309040

EM estimate for Wishart scale matrices:

fit4$Sigma
#> [[1]]
#>           [,1]     [,2]
#> [1,]  4.048859 -1.10103
#> [2,] -1.101030  2.79641
#> 
#> [[2]]
#>           [,1]      [,2]
#> [1,] 0.5582012 0.2515063
#> [2,] 0.2515063 0.8151752
#> 
#> [[3]]
#>           [,1]      [,2]
#> [1,] 2.0930529 0.6273737
#> [2,] 0.6273737 1.5706730

Reference

Mai TT, Zhao Z (2026). Mixture-of-experts Wishart model for covariance matrices with an application to Cancer drug screening. arXiv:2602.13888.