I'm updating cluster assignments in the context of a non-parametric Bayesian mixture model. When computing the probability of starting a new cluster, in the absence of cluster parameters (and using a conjugate model), we would evaluate the likelihood integrating over all cluster parameters.

In my particular case, I have data points $x$ and covariates $y$, which I model as follows: $$x_i \vert \Theta, z_i \sim \text{Normal}(\theta_{z_i} \cdot y_i, \sigma^2)$$ where $x_i$ is a datum $i$, $y_i$ are covariates, $z_i$ the cluster assignment, $\Theta$ all the cluster parameters. The mean for a data point is the dot product of cluster parameters and the point's covariates.

The prior over cluster parameters is multivariate normal: $$\theta \sim \text{Multivariate-Normal}(\mu, \Sigma)$$

What I wish to compute basically is $$P(z_i = z \vert \dots) \propto P(z) \int \!\! P(x_i \vert \theta)p(\theta)d\theta$$ where $z$ represents a new cluster.

Is there any hope to get an analytic solution? If not, how would you recommend to approach this? Monte Carlo integration?



Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge that you have read and understand our privacy policy and code of conduct.