一、主 题：Correlated samples from Dirichlet Process Marginal
Drawing bivariate samples from two Dirichlet Processes (DPs) can be a challenging yet important problem: in a setting where we wish to leave the individual stick-breaking weights marginal invariant, adding a copula function can be a plausible approach. However, the vanilla Gibbs sampling to the model can result in slow-mixing, yet, there is no analytical solution to a collapsed Gibbs Sampling where both the stick-breaking weights and the copula variable can be integrated out simultaneously. In this talk, we provide the details (and visualisation) of a partial collapsed approach: we integrate out either the stick-breaking weights or the copula variable condition on the other. Luckily, both conditional probabilities are in their close-form. We applied this framework in Mixed-Membership Stochastic Blockmodel where intra-group correlation can be modelled through the copula functions. This paper was first published in arXiv in June 2013 and then International Joint Conference on AI (IJCAI) in June 2016.
Yida Xu is the director of Machine Learning and Data Analytics Lab @GDBTC::UTS. He lead a group of 12 talented PhD students and engineers to apply our research and engineering skills to both Government and Retail insight analytics. In addition to cutting edge research, his group has also mastered modern data science tools, including Spark and TensorFlow.
Yida Xu is a world-renowned first line researcher in Machine Learning, Data Analytics, Computer Vision and Deep Learning. He wrote a series of Statistics, Probability and Machine Learning course for PhD students around the world. Research Interests are machine Learning, Data Analytics and Computer Vision.
编辑：林坤 / 审核：林坤 / 发布者：林坤