Mixture models
Context
In statistics (or machine learning, or other data modeling), distributions are usually about modelling a population.
The ability to model the entire population as a single distribution (gaussian, or whatever distribution may be more applicable to the case) into is a simplification that is both descriptive and has predictive use.
Now consider cases where there may be multiple contributions into the same data.
For a stolen example[1], the price of books might be bimodal: paperbacks around ten bucks, hardcovers around twenty. A model that asks to fitting a single gaussian would not summarize or predict very well. A model that asks to fit two would work measurably better.
A mixture model (sometimes mixture distribution) is a density model consisting of a mixture (weighed sum) of independent variables.
Exactly how abstract the idea is varies, because you could choose
- other amounts of groups (potentially infinite, but in practice often a relatively small number),
- the distribution used for each part (often a gaussian, in which case we're talking about a GMM)
- other types of data (not just real-numbered),
- more-dimensional data,
- more complex models.(verify)
So potentially quite fancy, though in practice (an in examples) usually
about modeling a small amount of distrinct contributions,
and often with an amount that we have a good reason to choose.
See also
- http://en.wikipedia.org/wiki/Gaussian_mixture_model
- http://www.csse.monash.edu.au/~dld/mixturemodel.html