Tuesday, May 20, 2014

no i dunno what its about, but i sure hope you do

http://ift.tt/1mRPerh



This article is the third part of the series on Clustering with Dirichlet Process Mixture Models. The previous time we defined the Finite Mixture Model based on Dirichlet Distribution and we posed questions on how we can make this particular model infinite. We briefly discussed the idea of taking the limit of the model when the k number of clusters tends to infinity but as we stressed the existence of such an object is not trivial (in other words, how do we actually “take the limit of a model”?). As a reminder, the reason why we want to take make k infinite is because in this way we will have a non-parametric model which does not require us to predefine the total number of clusters within the data.


Even though our target is to build a model which is capable of performing clustering on datasets, before that we must discuss about Dirichlet Processes. We will provide both the strict mathematical definitions and the more intuitive explanations of DP and we will discuss ways to construct the process. Those constructions/representations can be seen as a way to find occurrences of Dirichlet Process in “real life”.


Despite the fact that I tried to adapt my research report in such a way so that these blog posts are easier to follow, it is still important to define the necessary mathematical tools and distributions before we jump into using the models. Dirichlet Process models are a topic of active research, but they do require having a good understanding of Statistics and Stochastic Processes before using them. Another problem is that as we will see in this article, Dirichlet Processes can be represented/constructed with numerous ways. As a result several academic papers use completely different notation/conventions and examine the problem from different points of view. In this post I try to explain them as simple as possible and use the same notation. Hopefully things will become clearer with the two upcoming articles which focus on the definition of Dirichlet Process Mixture Models and on how to actually use them to perform cluster analysis






from Lizard's Ghost http://ift.tt/1teEw01

No comments:

Post a comment