In nicolasJouvin/MMPCA: Inference and Clustering for Mixture of Multinomial Principal Component Analysis

set.seed(42)

knitr::opts_chunk$set(
  collapse = TRUE,
  comment = "#>"
)

library(MoMPCA)
library(aricode)

Description

MMPCA is a package to perform clustering of count data based on the mixture of multinomial PCA model. It integrates a dimension reduction aspect by factorizing the multinomial parameters in a latent space, like Latent Dirichlet Allocation of Blei et. al. It specially conceived for low sample high-dimensional data. Due to the intensive nature of the greedy algorithm, it is not suited for large sample size.

Dataset

The package contains attached data in BBCmsg. It consists in 4 text document already preprocessed with the tm package. It is mostly useful for the simulate_BBC() function.

data("BBCmsg")

Demonstration for document clustering

Start by generating data from the MMPCA model with a particular $\theta^\star$ and $\beta^\star$. For more detail, check experimental section of the paper.

N = 200
L = 250
simu <- simulate_BBC(N, L, epsilon = 0, lambda = 1)
Ytruth <- simu$Ytruth

Then perform clustering

t0 <- system.time(res <- mmpca_clust(simu$dtm.full, Q = 6, K = 4,
                          Yinit = 'random',
                          method = 'BBCVEM',
                          max.epochs = 7,
                          keep = 1,
                          verbose = 2,
                          nruns = 2,
                          mc.cores = 2)
               )
print(t0)

Results analysis

tab <- knitr::kable(table(res@clustering, Ytruth), format = 'markdown')
print(tab)
cat('Final ARI is ', aricode::ARI(res@clustering, Ytruth))

Other visualization are also accessible from the plot function. Which takes several arguments

cond = requireNamespace("ggplot2", quietly = TRUE) & requireNamespace("dplyr", quietly = TRUE) & requireNamespace("tidytext", quietly = TRUE)

ggtopics <- plot(res, type = 'topics')
print(ggtopics)

ggbound <- plot(res, type = 'bound')
print(ggbound)

Model selection

The package contains a convenient wrapper around mmpca_clust() which performs model selection over a grid of values for $(K,Q)$. Here is the results for Qs = 5:7 and Ks = 3:5.

t1 <- system.time(res <- mmpca_clust_modelselect(simu$dtm.full, Qs = 5:7, Ks = 3:5,
                          Yinit = 'kmeans_lda',
                          init.beta = 'lda',
                          method = 'BBCVEM',
                          max.epochs = 7,
                          nruns = 3,
                          verbose = 1)
               )
print(t1)
best_model = res$models
print(best_model)

nicolasJouvin/MMPCA documentation built on Jan. 23, 2021, 3 a.m.