Part C/MSc
Bayes Methods
Mock 3 2017
1. (a) Let S = (S1, ..., SK ) be a partition of 1 : n = {1, 2, ..., n}.
(i) Specify the Chinese Restaurant process for arrivals 1 , 2, ..., n.
(ii) Let PCRP(S) equal the probability the outcome of the CRP is the random partition S. Show that
[Note that Γ(α + n) = Γ(α) (α + i - 1) .]
(iii) Show that PCRP(S) does not depend on the order of the arrivals.
(iv) Let i1 , i2 , i3 ∈ 1 : n be three fixed labels. What is the probability that i1 , i2 , i3 are in the same partition set?
(b) Consider a mixture of normal densities with a fixed number M of components, Dirichlet distributed mixture component weights w = (w1 , ..., wM ), and a prior π(θ*) for the
mixture component parameters θ* = (θ1(*), ..., θM(*)):
w ~ Dirichlet(α1, ..., αM) with αM = α/M for α > 0 fixed;
zi ~ Multinom(w), i = 1, ..., n;
θm(*) ~ π(θm(*)), m = 1, ..., M.
Here zi ~ Multinom(w), i = 1, ..., n means that zi = m, m ∈ {1, ..., M} with probability wm. In this model zi ∈ {1, ..., M} is the label of the cluster to which yi belongs. The observation model is
yi ~ f(yi; θz(*)i ), i = 1, ..., n.
Suppose the list z = (z1 , ..., zn) of cluster labels contains K ≤ M unique values m1 , ...mK . For k = 1, ..., K let Sk = {i : zi = mk, i = 1, ..., n}. Let S = (S1, ..., SK ). We write S = S(z) for the partition determined from z in this way.
(i) Write down the posterior for θ* , z, w|y in terms of the model elements.
(ii) Calculate the marginal prior probability πz(z) = ∫ πzjw(z|w)πW (w)dw for a set of cluster labels.
(iii) For k = 1, ..., K, let nk = |Sk| give the number of items in cluster k. Let
denote the prior distribution over partitions. Show that
(iv) Show that the prior distribution over partitions converges to a CRP as M → ∞ .
(v) Write down the distribution to which the marginal posterior distribution of θ* , S|y converges in the limit as M → ∞ .
版权所有:编程辅导网 2021 All Rights Reserved 联系方式:QQ:99515681 微信:codinghelp 电子信箱:99515681@qq.com
免责声明:本站部分内容从网络整理而来,只供参考!如有版权问题可联系本站删除。