erte pan wireless eng. group advisor: dr. han department of electrical and computer engineering...
TRANSCRIPT
Erte PanWireless Eng. Group
Advisor: Dr. Han
Department of Electrical and Computer EngineeringUniversity of Houston, Houston, TX.
Erte PanWireless Eng. Group
Advisor: Dr. Han
Department of Electrical and Computer EngineeringUniversity of Houston, Houston, TX.
Mercer Kernel-Based Clustering in Feature Space
For Math6397 Prof. Azencott
Mercer Kernel-Based Clustering in Feature Space
For Math6397 Prof. Azencott
Author: Mark GirolamiSubmitted in IEEE Transactions on Neural
Netwroks, Vol.3, May, 2002Citations so far: 593
Author: Mark GirolamiSubmitted in IEEE Transactions on Neural
Netwroks, Vol.3, May, 2002Citations so far: 593
ContentContent
Problem Statement
Data-space Clustering
Feature-space Clustering
Stochastic Optimization
Nonparametric Clustering
Results and Discussion
References
Problem StatementProblem Statement
A lot of data analysis or machine learning tasks involve classification of data clouds or prediction of incoming data point.
Machine Learning: Enable computers to learn without being explicitly programmed.
Unsupervised Learning
Supervised Learning
Data-space ClusteringData-space Clustering
Clustering: Unsupervised partitioning of data observations into self-similar regions.
Traditional clustering method:
Centroid-based clustering
Hierarchical clustering
Distribution-based clustering…
Data-space ClusteringData-space Clustering
Problem formulation:
N data vectors in D-dimension space:
Given K cluster centers the within-cluster scatter matrix is defined as:
where the binary variable indicates the membership of data point to cluster k
Dnn xNnx ...,2,1,
K
k
N
n
TknknknW mxmxz
NS
1 1
))((1
knz
nx
N
nnkn
kk xzN
m1
1
N
nknk zN
1
Data-space ClusteringData-space Clustering
Data-space clustering criterion: sum-of-squares; measure of compactness
K-means, mean shift and so forth…
The partition of data set is solved by the optimization problem:
NP-hard problem… heuristic algorithms such as Lloyd’s algorithm:
Initialize centroid for given number of clusters k
Assign each data point to the “nearest” mean(Voronoi diagram)
Update centroids of each clusters
)Tr(minarg WZ
SZ
Data-space ClusteringData-space Clustering
Drawbacks of Data-space clustering:
linear separation boundaries.
prefer similar size of each cluster.
equally weighted in each dimension.
number of clusters, K, has to be determined at the beginning.
might stuck into a local minimum.
sensitive to initialization and outliers. Feature-space clustering is proposed to address those problems, hopefully…
Feature-space ClusteringFeature-space Clustering
Same story as in Kernel PCA that everyone can recite…
FD : Xx
Feature-space ClusteringFeature-space Clustering
Computation in feature space, utilizing the kernel trick:
}))()()((1
Tr{)Tr(1 1
K
k
N
n
TknknknW mxmxz
NS
K
k
N
nkn
TknknW mxmxz
NS
1 1
))(())((1
)Tr(
Using the Mercer Kernels, the Gram Matrix is: )(),(),( jijiijji xxxxkKK
Denote the term:
then:
N
i
N
lilklki
k
N
jnjkj
knnkn Kzz
NKz
NKy
1 12
1
12
K
k
N
nknknW yz
NS
1 1
1)Tr(
Feature-space ClusteringFeature-space Clustering
Denote the following terms:
Then the straightforward manipulation of the equations yield:
If the Radial Basis Function kernel is used:
Then the first term reduces to unity, thus:
captures the quadratic sum of the elements allocated to the k-th cluster
NNk
k ijkj
N
i
N
j kikk KzzNCxR
1 1
2)|(
)|(1
)Tr(11 1
k
K
kk
K
k
N
nnnknW CxRKz
NS
}||||)/1(exp{),( 2jiji xxcxxk
)|(1)Tr(1
k
K
kkW CxRS
)|( kCxR
Feature-space ClusteringFeature-space Clustering
For the RBF kernel, the following approximation hold due to the convolution theorem for Gaussians(why?):
This being the case, then:
It make sense for the clustering later on, because the integral is the measurement of the compactness of the cluster
Connectivity to Probability Statistics; Validation of the kernel model. (What if not RBF kernels? This proves they are not valid?)
N
i
N
jijxK
Ndxxp
1 12
2 1)(
dxCxpKzzN
CxRkCx
k
N
i
N
jijkjki
kk
2
1 12
)|(1
)|(
Feature-space ClusteringFeature-space Clustering
Make sense of the integral :
Utilizing the Cauchy’s Inequality in statistics:
The equality holds when , which means the more “uniformly” distributed data, the more compact cluster.
Examples:
Gaussians:
dxxpx
2)(
)}({)( 2 xpEdxxpx
}1{})({}1)({ 2 ExpExpE
1)( axp
21)( 2 dxxp
x
Feature-space ClusteringFeature-space Clustering
The integral represented by is the contrast to the Euclidean compactness measure defined by the sum-of-squares term.
Now the optimization problem in the feature-space becomes:
Lemma: If the binary restriction for is relaxed to , then the optimization above is achieved with Z matrix being binary.
Interpretation: the optimal partitioning of data will only occur when the partition indexes are 0 or 1.
This validates the use of stochastic methods in optimizing.
)|( kCxR
K
kkk
ZW
ZCxRSZ
1
)|(maxarg)Tr(minarg
knz 10 knz
Stochastic OptimizationStochastic Optimization
Define as the penalty associated with assigning the j-th data point to the k-th cluster in feature-space.
Due to the nature of RBF kernel, the range of each element of K would be (0,1].
The second term of the penalty can be viewed as estimate of the conditional probability of the j-th data given the k-th cluster.
The original objective of optimization problem is manipulated into:
N
ljlkl
kkj Kz
ND
1
11
}||||)/1(exp{),( 2jiji xxcxxk
N
j
K
k
N
ljl
k
klkjW K
N
zz
NS
1 1 1
11)Tr(
N
j
K
kkjkjW Dz
NS
1 1
1)Tr(
Stochastic OptimizationStochastic Optimization
Analog to the stochastic optimization in data-space:
where the Ekn is the sum-of-squares distance term.
Solved as the fashion of the Expectation Maximization algorithm:
The cluster indicator is calculated according to its expectation employing softmax function:
each is then updated by the newly estimated expectation values of the indicators
N
n
K
kknknW Ez
NS
1 1
1)Tr(
knz
K
k
newnk
newkn
kn
E
Ez
)exp(
)exp(
2|||| kknewkn mxE
knz
Stochastic OptimizationStochastic Optimization
Similarly, the stochastic optimization in feature-space:
where:
note that indicates the compactness of the k-th cluster.
K
k
newnkk
newknk
K
k
newnk
newkn
kn
D
D
y
yz
11
)2exp(
)2exp(
)exp(
)exp(
))|(exp( kk CxR
N
ljlkl
k
newkn Kz
ND
1
11
k
Stochastic SearchStochastic Search
Stochastic method for optimization
Different optimization criteria in traditional method and stochastic method for optimization purpose:
Traditional: Error criterion. BP method strictly goes along the gradient descent direction. Any direction that enlarge error is NOT acceptable. Easy to get stuck in local minima.
BM: associate the system with “Energy”. Simulated Annealing enables the energy to grow under certain probability.
Simulated AnnealingSimulated Annealing
Simulated Annealing:
1. Create initial solution Z (global states of the system)
Initialize temperature T>>1 2. Repeat until T =T-lower-bound
Repeat until thermal equilibrium is reached at the current T• Generate a random transition from Z
to Z’• Let E = E(Z’) E(Z)• if E < 0 then Z = Z’ • else if exp[E/T] > rand(0,1) then Z = Z’
Reduce temperature T according to the cooling schedule
3. Return Z
1. Create initial solution Z (global states of the system)
Initialize temperature T>>1 2. Repeat until T =T-lower-bound
Repeat until thermal equilibrium is reached at the current T• Generate a random transition from Z
to Z’• Let E = E(Z’) E(Z)• if E < 0 then Z = Z’ • else if exp[E/T] > rand(0,1) then Z = Z’
Reduce temperature T according to the cooling schedule
3. Return Z
This term allows “thermal disturbance” which facilitate finding global minimum
Nonparametric ClusteringNonparametric Clustering
Nonparametric: No assumptions on the number of clusters.
Observations:
the kernel matrix will have a block diagonal structure when there are definite clusters within the data.
eigenvectors of a permuted matrix are the permutations of the original matrix and therefore, an indication of the number of clusters may be given from the eigen-decomposition of kernel matrix.
Recall the approximation:
N
i
N
jijxK
Ndxxp
1 12
2 1)(
Nonparametric ClusteringNonparametric Clustering
Moreover,
Eigen-decomposition of K gives:
Thus we have:
This indicates that if there are K distinct clusters within the data samples then there will be K dominant terms in (Why?)
NTN
N
i
N
jijx
KKN
dxxp 111
)(1 1
22
TUUK
N
ii
TNiN
Tii
N
ii
TNN
TN uuuK
1
2
1
}1{1}{111
2}1{ iTNi u
Results and DiscussionResults and Discussion
Results on phantom 3 data sets: Fisher Iris; Wine data set; Crabs data.
Results and DiscussionResults and Discussion
Conclusions and discussions:
the mean vector in feature-space may not serve as representatives or prototypes of the input space clusters.
the block-diagonal structure of the kernel matrix can be exploited in estimating the number of possible clusters.
choice of kernel will be data specific.
the RBF kernels link the sum-of-squares criterion with the probability metric.
the choice of the parameter of RBF kernel should be determined by the cross-validation or the leave-one-out technique.
eigen-decomposition of N x N kernel matrix scales as O(N^3)
Results and DiscussionResults and Discussion
Remarks of my own:
most appealing point is the link between distance metric and the probability metric.
unclear about why prefer to use the stochastic optimizing instead of ordinary optimizing methods.
no assessment on other types of kernels.
unclear about how to permute the kernel matrix to get the block-diagonal structure.
the “super technical” term “dominant” in the non-parametric part is too vague; needs some quantification.
2}1{ iTNi u
ReferencesReferences
“Data clustering and data visualization”, in Learning in Graphical Models,1998.
“A projection pursuit algorithm for exploratory data analysis”, IEEE Trans. Comput., 1974.
“An algorithm for Euclidean sum-of-squares classification”, Biometrics, 1988
“Maximum certainty data partitioning”, Pattern Recognition, 2000.
“An expectation maximization approach to nonlinear component analysis”, Neural Comput., 2001