2
$\begingroup$

So I'm trying out a toy problem of inferring the mixing weights of a K-component Gaussian mixture model (just the weights, so I'm assuming the parameters of each Gaussian is known). My posterior is $\pi({\bf p}|\{x_i\})\propto[\prod_{i=1}^{n}\sum_{k=1}^{K}\pi_k(x_i)p_k][\prod_{k=1}^K p_k^{\alpha_k-1}]$, where $\pi_k$ is a Gaussian density, and there's a Dirichlet prior on the mixing weights.

At first I tried M-H using a Dirichlet proposal distribution with parameters equal to the previous sample of $\bf p$ scaled by some factor (i.e. ${\bf p_{prop}}\sim Dir({\bf p_{curr}}\times scaling\ factor) $), but basically none of the samples would be accepted unless my prior for $\bf p$ was close to the true distribution (e.g. if the true weights were 0.5, 0.25, 0.25 and my prior was $Dir(50,25,25)$). Does anyone have any insight into why a Dirichlet proposal doesn't work?

I then found these slides (https://www.ceremade.dauphine.fr/~xian/BCS/Bmix.pdf) which say to reparameterize $\bf p$ by a vector $\bf w$ where $p_i=\frac{w_i}{\sum_{j=1}^K w_j}$, so that you can sample the $\bf w$ without the constraint of them adding to 1. I then used the following M-H step:

-propose $w_1'\sim Norm(w_1,\sigma^2)$ (automatically reject the sample if $w_1<0$)

-test the current sample of $\bf w$ with $w_1$ replaced by $w_1'$, if accepted replace $w_1$ with $w_1'$ in the current sample

$\vdots$

-propose $w_k'\sim Norm(w_k,\sigma^2)$ (automatically reject the sample if $w_k<0$)

-test the current sample of $\bf w$ with $w_k$ replaced by $w_k'$, if accepted replace $w_k$ with $w_k'$ in the current sample

-store $\bf w$ as the sample for the current time step

Is this a valid way to sample $\bf w$, or should I be only testing out the proposed sample after sampling every component (i.e. propose ${\bf w'}\sim Norm({\bf w},\sigma^2 I)$ and then test that sample)? Or should I be doing something else altogether?

$\endgroup$
3
$\begingroup$

At first I tried M-H using a Dirichlet proposal distribution with parameters equal to the previous sample of p scaled by some factor (i.e. $p^\text{prop}\sim\text{Dir}(p^\text{curr}×\alpha)$ where $\alpha$ is a scaling factor), but none of the samples would be accepted unless my prior for p was close to the true distribution. Does anyone have any insight into why a Dirichlet proposal doesn't work?

This should work when adapting the scaling factor $\alpha$ to be rather small and possibly adding a stabilising term $\delta$ as in $p^\text{prop}\sim\text{Dir}(p^\text{curr}×\alpha+\delta)$. And of course using the right Metropolis-Hastings acceptance ratio, since the proposal is not a random walk.

Is this a valid way to sample w, or should I be only testing out the proposed sample after sampling every component (i.e. propose $w′∼Norm(w,σ^2I)$ and then test that sample)? Or should I be doing something else altogether?

Simulating the whole vector $\mathbf{w}$ at once or one component of $\mathbf{w}$ at a time are both valid approaches (provided one uses the proper Metropolis-Hastings acceptance ratio, obviously). Acceptance rates should be higher for the unidimensional case but convergence slower.

$\endgroup$
3
$\begingroup$

Or should I be doing something else altogether?

There is a different approach that often works quite well. It involves introducing latent classification variables: Each observation is classified with a mixture component, but these classifications are not observed.

By way of introduction, consider generating random variates from a mixture distribution. One way is to first use the mixture weights (they characterize a categorical distribution) to choose the mixture component (i.e., choose the classification) and then draw the variate from that component.

Drawing from the posterior distribution is fairly straightforward. The combination of the mixture weights and the classifications forms the basis of a Gibbs sampler. Let $z = (z_1, \ldots, z_n)$ where $z_i = k$ indicates that observation $i$ belongs to component $k$. Let $(p^{(r)},z^{(r)})$ denote the current state of the sampler. Let $c = (c_1,\ldots,c_K)$, where $c_k$ denotes the multiplicity of $k$ in $z$ (i.e., the number of observations classified with component $k$). The draw of the mixture weights is a draw from the Dirichlet distribution (as updated by the classifications): \begin{equation} p^{(r+1)}\,|\,z^{(r)} \sim \textsf{Dir}(\alpha + c^{(r)}) . \end{equation} Given the mixture weights and the observations, each classification is drawn from a categorical distribution where \begin{equation} (z_i^{(r+1)} = k)\,|\,(p^{(r+1)},x)\ \propto\ p_k^{(r+1)}\,\pi_k(x_i) \qquad\text{for $k = 1, \ldots, K$}. \end{equation} Note that $p_k^{(r+1)}$ plays the role of the prior for component $k$ while $\pi_k(x_i)$ plays the role of the likelihood.

$\endgroup$
  • $\begingroup$ Reading this made me work through deriving the conditionals which I was hesitant to do (was confused how exactly to do it before), thank you! Only thing is the sampling of $p^{(r+1)}$ should be a Dirichlet where you're adding the counts of the $z^{(r)}$ for each component to $\alpha$, not $z^{(r)}$ itself. $\endgroup$ – aleshing Sep 5 '17 at 19:52
  • $\begingroup$ You are quite right about the counts. I'll fix that. Glad my answer was helpful. $\endgroup$ – mef Sep 5 '17 at 21:28

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.