Skip to content

Latest commit

 

History

History
66 lines (44 loc) · 2.81 KB

3-Mixture-model-applications.md

File metadata and controls

66 lines (44 loc) · 2.81 KB

Mixture model applications

Mixture models are a very general and useful tool. Here are a few scientific papers that utilize mixture models in a variety of contexts from Bayesian inference to distribution function modeling of stellar orbits.

Data-driven models:

  • Bovy et al. 2010

    They use extreme deconvolution (XD) to build a star/quasar classifier to be used for spectroscopic target selection for the SDSS.

  • Anderson et al. 2017

    We used XD to infer the density of sources in the color-magnitude diagram, then used the inferred density to improve distance estimates for faint/distant stars in Gaia DR1.

  • Melchior & Goulding 2016

    This is the paper describing the method Peter spoke about a few weeks ago: XD with support for incomplete data. They apply it to the problem of deconvolving an X-ray (Chandra) image of a nearby galaxy that has masked regions, chip gaps, and known sensitivity variations across the detector.

  • Hao et al. 2009

    Similar to XD, but coming from a pure probability weighted approach it lacks the marginalization over the not-observed true sample positions. This should lead to and over-interpretation of the data.

Mixture-models in likelihoods:

  • Hogg et al. 2010

    Section 3 describes how to fit a model in the presence of outliers in a probabilistic sense. The likelihood simplifies to being a mixture model. See also Dan F-M's blog post.

  • Price-Whelan et al. 2015

    This is just an example application of the outlier model from Hogg et al. 2010.

  • Kelly 2007

    Similar treatment to XD for regression with errors. Includies selection effects and non-detections. Gibbs sampler.

Mixture-models for convenience:

  • Oh et al. 2017

    Used a Gaussian mixture model to represent the true distribution of velocities in the Milky Way disk, i.e. as a prior. In principle, we could have used any distribution with tails, but using a GMM with a Gaussian likelihood meant that many marginalizations are analytic.

  • Hogg & Lang 2012 and Sheldon 2014

    Replace the Sersic model for galaxies with a mixture of Gaussians, so that convolution with the PSF (also represented as mixture of Gaussian) becomes analytic and fast.

  • Also ask Scott Carlsten about his project!

Flexible theoretical models: