Transcript Lecture30

Nanjing University of Science & Technology
Pattern Recognition:
Statistical and Neural
Lonnie C. Ludeman
Lecture 30
Nov 11, 2005
1
Lecture 30 Topics
1. General Comments about the
Clustering Problem
2. Present my small programs that can
be used for performing clustering
3. Demonstrate the programs
4. Closing Comments
2
Clustering is the art of grouping
together pattern vectors that in some
sense belong together because they
have similar characteristics and are
different from other pattern vectors.
In the most general problem the
number of clusters or subgroups is
unknown as are the properties that
make them similar.
Review
3
Question:
How do we start the process of finding
clusters and identifying similarities???
Answer:
First realize that clustering is an art and
there is no correct answer only feasible
alternatives.
Second explore structures of data,
similarity measures, and limitations of
various clustering procedures
Review
4
Problems in performing meaningful clustering
Scaling
The nonuniqueness of results
Programs always give clusters
even when there are no clusters
Review
5
There are no correct answers, the
clusters provide us with different
interpretations of the data where the
closeness of patterns is measured with
different definitions of similarity.
The results may produce ways of looking
at the data that we have not considered or
noticed. These structural insights may
prove useful in the pattern recognition
process.
Review
6
Methods for Clustering Quantitative Data
1. K-Means Clustering Algorithm
2. Hierarchical Clustering Algorithm
3. ISODATA Clustering Algorithm
4. Fuzzy Clustering Algorithm
Review
7
K-Means Clustering Algorithm
Randomly Select K cluster centers from
Pattern Space
Distribute set of patterns to the cluster
center using minimum distance
Compute new Cluster centers for each
cluster
Continue this process until the cluster
centers do not change.
Review
8
Agglomerative Hierarchical Clustering
Consider a set S of patterns to be clustered
S = { x1, x2, ... , xk, ... , xN}
Define Level N by
= { x1}
(N)
S2
= { x2}
(N)
SN
Review
...
S1
(N)
= { xN}
Clusters at
level N are the
individual
pattern vectors
9
Define Level N -1 to be N – 1 Clusters
formed by merging two of the Level N
clusters by the following process.
Compute the distances between all the
clusters at level N and merge the two with
the smallest distance (resolve ties
randomly) to give the Level N-1 clusters as
(N-1)
S1
(N-1)
S2
...
Review
SN-1
(N-1)
Clusters at
level N -1 result
from this
merging
10
The process of merging two clusters at
each step is performed sequentially until
Level 1 is reached. Level one is a single
cluster containing all samples
(1)
S1
= { x1, x2, ... , xk, ... , xN}
Thus Hierarchical clustering provides
cluster assignments for all numbers of
clusters from N to 1.
Review
11
Fuzzy C-Means Clustering Preliminary
Given a set S composed of pattern
vectors which we wish to cluster
S = { x1, x2, ... , xN}
Define C Cluster Membership Functions
1
2
2
2
C
Review
C
...
1
...
1
C
C
)]
12
Define C Cluster Centroids as follows
Let Vi be the Cluster Centroid for Fuzzy
Cluster Cli , i = 1, 2, …, C
Define a Performance Objective Jm as
where
Review
13
Definitions
A is a symmetric positive definite matrix
Ns is total number of pattern vectors
m = Fuzziness Index (m >1 )
Higher numbers being more fuzzy
The Fuzzy C-Means Algorithm minimizes Jm
by selecting Vi and i , i =1, 2, … , C by an
alternating iterative procedure as described in
the algorithm’s details
Review
14
Fuzzy C-Means Clustering Algorithm
(a) Flow Diagram
No
Review
Yes
15
General Programs for Performing Clustering
1. Available commercial Packages:
SPSS , SAS, GPSS,
2. Small Programs for classroom use
LCLKmean.exe
LCLHier.exe
LCLFuzz.exe
16
2. Small Programs for classroom use
LCLKmean.exe
Use the K-Means Algorithm to cluster
small data sets
LCLHier.exe
Performs Hierarchical Clustering
of small data sets
LCLFuzz.exe
Performs Fuzzy and crisp clustering
of small data sets
17
Data File Format for the LCL Programs
NS = Number of data samples
VS = Data vector size
DATA in row vectors with space
between components
NS
VS
DATA
5
3
163
205
714
668
223
Text
File
18
Food for Thought
All the clustering techniques presented so
far use a measure of distance or similarity.
Many of these give equal distance contours
that represent hyper spheres and hyper
ellipses.
If these techniques are used directly on
patterns that are not describable by those
type of regions we can expect to obtain
poor results.
19
In some cases each cluster occupies a limited
region (subspace of the total pattern space )
described by a nonlinear functional relation
between components. An example appears below.
Existing
Pattern
vectors
Existing
Pattern
Vectors
Standard K-Means, Hierarchical, or Fuzzy
cluster analysis directly on the data will produce
unsatisfactory results.
20
For this type of problem the patterns
should be first preprocessed before a
clustering procedure is performed .
Two almost contradictory approaches
can be used for this processing.
1. Extend the pattern space by techniques
comparable to functional link nets so that the
clusters can be separated by spherical and
elliptical regions.
2. Reduce the dimension of the space by
a nonlinear form of processing involving
principal component like processing before
clustering.
21
Both methods imply that we know
additional information about the structure
of the data.
This additional information may be known to
us or it may need to be determined.
The process of finding structure within data
has been put in the large category of
“Data Mining”.
So get a shovel and start looking. Good luck
in your search for gold in the mounds of
practical data.
22
Several very important topics in Pattern
Recognition were not covered in this course
because of time limitations. The following topics
deserve your special attention to make your
educational experience complete
1. Feature Selection and Extraction
2. Hopfield and feedback neural nets
3. Syntactical Pattern Recognition
4. Special Learning Theory
23
Like to Thank
Nanjing University of Science & Technology
and
Lu Jian Feng
Yang Jing-yu
Wang Han
for inviting me to present this course
on
Statistical and Neural Pattern Recognition
24
A Very Special Thanks to my new friends
Lu Jian Feng
Wang Qiong
Wang Huan
for looking after me. Their kindness
and gentle assistance has made my
stay in Nanjing a very enjoyable and
unforgettable experience.
25
Last and not least I would like to thank
all you students
for your kind attention throughout this
course. Without your interest and cheerful
faces it would have been difficult for me to
teach.
My apology for teaching in English, which
I am sure, made your work a lot harder.
Best of Luck to all of you in your studies
and life.
26
“As you travel through life may all
your trails be down hill and the wind
always be at your back”.
Bye for now and I hope our paths
cross again in the future.
I will have pleasant thoughts about
NUST Sudents and Faculty, Nanjing,
and China as I head back to New
Mexico !
27
New Mexico
Land of Enchantment
28
End of Lecture 30
29