4
$\begingroup$

I'm using the Mclust function of the mclust package in R to fit a mixture of Gaussians model. My simulated data obviously has 3 components:

# Example data
set.seed(101)
data = c(rnorm(100, mean = 10), rnorm(20, mean = 20), rnorm(50, mean = 15))
hist(data)

enter image description here

However, when I learn the parameters of the fitted gaussian mixture using Mclust:

require(mclust)
# Run Mclust
mixmdl = Mclust(data)
# Show means of fitted gaussians
print( mixmdl$parameters$mean ) 

Why are there two Gaussians fitted under the first mixture. Is there an option I can set to avoid overfitting like this?

$\endgroup$
  • $\begingroup$ The problem most likely is one of initialization. Have you tried other means of initializing Mclust? $\endgroup$ – Anony-Mousse Jan 27 '15 at 22:25
4
$\begingroup$

I think this isn't that big of a deal. Any statistical procedure, from calculating a mean to fitting a GLMM with crossed random effects, is a function of your data. And your data are, by assumption, a mix of signal and stochastic noise. Thus your output is random as well. In the long run, you hope it will be centered on the true value (unbiased, consistent), and in the short run you hope that the output will be similar / close to the true picture. Nothing about this changes when we move to talking about Gaussian mixture models.

For what it's worth, the model seems decent to me. Here's the standard output that I get for your situation:

summary(mixmdl)
# ----------------------------------------------------
# Gaussian finite mixture model fitted by EM algorithm 
# ----------------------------------------------------
#   
# Mclust E (univariate, equal variance) model with 4 components:
#   
#  log.likelihood   n df       BIC       ICL
#       -391.8016 170  8 -824.6895 -955.0467
# 
# Clustering table:
#  1  2  3  4 
# 12 89 49 20 

enter image description here

The BIC is optimized for your data by a mixture of four components with equal variances. You actually have three components, but look at the density plot in the lower right hand corner—it's not that far from the truth.

Consider the following plot. The solid black line is the density from the fitted Gaussian mixture model, the dashed red line is a naive kernel density estimate, and the dotted blue line is the underlying data generating process. It looks to me like Mclust did a good job of recovering something very close to the true DGP from your data.

enter image description here

Code:

xs = seq(7, 22, by=.1)
windows()
  plot(mixmdl, what="density")
  lines(density(data), col="red", lty=2)
  lines(x=xs, y=10/17*dnorm(xs, mean=10, sd=1) +
                 5/17*dnorm(xs, mean=15, sd=1) +
                 2/17*dnorm(xs, mean=20, sd=1), 
        col="blue", lty=3)
  legend("topright", legend=c("GMM","KDE","DGP"), col=c("black","red","blue"), lty=1:3)
$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.