Transcript Hierarchical Beta Process and the Indian Buffet Process

Hierarchical Beta Process and
the Indian Buffet Process
by R. Thibaux and M. I. Jordan
Discussion led by
Qi An
Outline
•
•
•
•
•
•
•
Introduction
Indian buffet process (IBP)
Beta process (BP)
Connections between IBP and BP
Hierarchical beta process (hBP)
Application to document classification
Conclusions
Introduction
• Mixture models
VS.
– Each data is drawn from
one mixture component
– Number of mixture
components is not set a
prior
– Distribution over partitions
• Factorial models
– Each data is associated
with a set of latent
Bernoulli variables
– Cardinality of the set of
features can vary
– A “featural” description of
objects
– A natural way to define
interesting topologies on
cluster
– May be appropriate for
large number of clusters
Beta process
Beta process is a special case of independent increment process, or Levy process,
Levy process can be characterized by Levy measure. For beta process, it is
If we draw a set of points (wi , pi )   [0,1] from a Poisson process with base
measure v, then
As the representation shows, B is discrete with probability one.
When the base measure B0 is discrete: B0   qi w , then B has atoms at the same
i
locations B   pi w with pi ~ Beta(cqi , c(1  qi ))
i
i
i
Bernoulli process
Here, Ω can be viewed as a set of potential features and the random measure
B defines the probability that X can possess particular feature.
In Indian buffet process, X is the customer and its features are the dishes the
customer taste.
Connections between IBP and BP
It is proven that the observations from a beta process satisfy
Procedure:
The first customer will try Poi(γ) number of dishes (feature).
After that , the new observation can taste previous dish j with
probability
and then try a
number of new features
where
is the total mass
As a result, beta process is a two-parameter (c, γ) generalization of
the Indian buffet process.
IBP=BP(c=1, γ=α)
The total number of unique dishes can be roughly represented as
This quantity becomes Poi(γ) if c0 (all customers share the same dishes)
or Poi(n γ) if c∞ (no sharing).
An algorithm to generate beta
process
Authors propose to generate an approximation,
Let
For each step n≥1
, of B
Hierarchical beta process
Consider a document classification problem. We have a training data set X,
which is a list of documents. Each document is classified by one of n topics. We
model a document by the set of words it contains. We assume document Xi,j is
generated by including each word w independently with a probability pjw specific
to topic j. These probabilities form a discrete measure Aj over all word space Ω.
We can put a beta process BP(cj,B) prior on Aj.
Since we want the sharing across different topics, B has to be discrete. We thus
put a beta process prior BP(c0,B0) on B, which allows sharing the same atoms
among topics.
The HBP model can be summarized as:
This model can be solved with Monte Carlo inference algorithm.
Applications
• Authors applied the hierarchical beta process to
a document classification problem
• Compare it to the Naïve Bayes (with Laplace
smoothing) results
• The hBP model can obtain 58% result while the
best Naïve Bayes result is 50%
Conclusions
• The beta process is shown to be suitable for
nonparametric Bayesian factorial modeling
• The beta process can be extended to a
recursively-defined hierarchy of beta process
• Compared to the Dirichlet process, the beta
process has the potential advantage of being an
independent increments process
• More work on inference algorithm is necessary
to fully exploit beta process models.