Scribd
Upload
27 views

Mixture Models and EM Algorithm: S. Sumitra

The document summarizes mixture models and the Expectation-Maximization (EM) algorithm. It explains that mixture models represent cluster distributions as a combination of component distributions like Gaussians. The EM algorithm is used to determine the unknown parameters of a mixture model, including component weights, means, and covariances. It does this through an iterative E-step to calculate membership probabilities, and an M-step to re-estimate the parameters based on these probabilities.

Uploaded by

kadarsh226521
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
27 views

Mixture Models and EM Algorithm: S. Sumitra

The document summarizes mixture models and the Expectation-Maximization (EM) algorithm. It explains that mixture models represent cluster distributions as a combination of component distributions like Gaussians. The EM algorithm is used to determine the unknown parameters of a mixture model, including component weights, means, and covariances. It does this through an iterative E-step to calculate membership probabilities, and an M-step to re-estimate the parameters based on these probabilities.

Uploaded by

kadarsh226521
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 4

Mixture Models and EM Algorithm

S. Sumitra

Clustering problems could be solved by applying model-based approach, which


consists in using certain models for clusters and attempting to optimize the fit be-
tween the data and the model. Each cluster (component) can be mathematically
represented by a parametric distribution, for eg, Gaussian (continuous) or a poisson
(discrete). The entire data set is therefore modelled by a mixture of these distribu-
tions. An individual distribution used to model a specific cluster is often referred to
as a component distribution.
Let there be k clusters. Let the random variable C denote the component with
valuesP1, ..k. Here we are considering Gaussian mixture models. So xj /(C = i) ∼
N (µi , i ) where µi and Σi are the mean and covariance matrix of the ith class.
A data point is generated by first choosing a component and then generating a
sample from that component. By total probability theorem,
k
X
p(x) = p(C = i)p(x/C = i) (1)
i=1

[p(C = i) is analogous to p(y = i) in Gaussian discriminant analysis.]


To determine in which cluster each xj belongs, p(C = i/xj ) has to be found. Now

p(C = i)p(xj /C = i)
p(C = i/xj ) = pij = , i = 1, 2, . . . k, j = 1, 2, . . . N (2)
p(xj )

Hence ki=1 pij = 1. Let wi = p(C = i), i = 1, 2. . . . k. Therefore the unknown


P
parameters of a mixture of Gaussians are wi , µi and Σi .
The EM algorithm can be applied to determine the unknown parameters. The
EM algorithm has two main steps: E step & M step. In the E-step, it assumes
the values of the model (that is wi , µi and Σi ) and find P (C = i/xj ), i = 1, 2, . . . k,
j = 1, 2, . . . N . In the M-step, it updates the parameters of the model. The process
iterates until convergence.

1
E step
In the E step, compute the probabilities pij , i = 1, 2, . . . k, j = 1, 2, . . . N.

M step
Compute the new mean, covariance and component weights as follows:
PN
j=1 pij xj
µi = PN
j=1 pij
P
j 1{xj ∈ C = i}xj
[For sure event, µi = P . Here, we don’t know whether xj is in
j 1.{xj ∈ C = i}
component i. We only know p(C = i/xj ).]
T
P
j pij (xj − µi )(xj − µi )
Σi = PN
j=1 pij
PN
j=1 pij
wi =
N
[Compare these formulas with those of Gaussian discriminant analysis]
The algorithm can be summarized as follows:

2
Algorithm 1 EM algorithm
Initialize µi , Σi , wi , i = 1, 2, . . . k
Iterate until covergence:
E Step
for i = 1 to k do
for j = 1 to N do
1 1
calculate p(xj /C = i) = n/2 1/2
exp − (xj − µi )T Σ−1
i (xj − µi )
(2π) |Σi | 2
PN 
calculate pij = (p(xj /C = i)wi ) / j=1 p(x j /C = i)w i
end for
pi = N
P
j=1 pij
end for
M Step
for i = 1 to k doP
N
j=1 pij xj
calculate µi =
PN p i T
j=1 pij (xj − µi )(xj − µi )
calculate Σi =
pi
pi
set wi =
N
end for
end

3
References
(1) Artificial Intelligence by Stuart Russel and Peter Norwig
(2) Andrew Ng’s Lecture Note

You might also like