Advisor : Dr. Hsu Presenter : Ai-Chen Liao Authors : Yiu-ming Cheung

1Intelligent Database Systems Lab

國立雲林科技大學National Yunlin University of Science and Technology

On Rival Penalization Controlled Competitive Learning for Clustering with

Automatic Cluster Number Selection

Advisor : Dr. Hsu

Presenter : Ai-Chen Liao

Authors : Yiu-ming Cheung

2005 . TKDE . Page(s) : 1583 - 1588

2

Intelligent Database Systems Lab

N.Y.U.S.T.

I. M.Outline

Motivation Objective Method

RPCL RPCCL

Experimental Results Conclusion Personal Opinions

3

Intelligent Database Systems Lab

N.Y.U.S.T.

I. M.

Motivation

K-means algorithm has at least two major drawbacks:─ It suffers from the dead-unit problem.─ If the number of clusters is misspecified, i.e., k is not equal

to the true cluster number k*, the performance of k-means algorithm deteriorates rapidly.

The performance of RPCL is sensitive to the value of the delearning rate.

4

Intelligent Database Systems Lab

N.Y.U.S.T.

I. M.Objective

We will concentrate on studying the RPCL algorithm and propose a novel technique to circumvent the selection of the delearning rate.

We further investigate the RPCL and present a mechanism to control the strength of rival penalization dynamically.

5

Intelligent Database Systems Lab

N.Y.U.S.T.

I. M.Method ─ RPCL

Advantage :

RPCL can automatically select the correct cluster number by gradually driving redundant seed points far away from the input dense regions.

Drawback :

RPCL is sensitive to the delearning rate. Idea :

ex. In a election campaign…..(more intense)…..

candidates : A 40%

B 35%

C 5%

6

Intelligent Database Systems Lab

N.Y.U.S.T.

I. M.Method ─ RPCL

cluster centereach input

Winner (move closer)

Rival (move away)

unchanged

7

Intelligent Database Systems Lab

N.Y.U.S.T.

I. M.Method ─ RPCL

8

Intelligent Database Systems Lab

N.Y.U.S.T.

I. M.Method ─ RPCCL

9

Intelligent Database Systems Lab

N.Y.U.S.T.

I. M.Method ─ RPCCL

This penalization control mechanism by

with

compare

10

Intelligent Database Systems Lab

N.Y.U.S.T.

I. M.Experimental ResultsRPCL : learning rate αC at 0.001, and αr at 0.0001the number of seed points : 30

audience image : 128*128 pixels

epoch :50

original Audience Image RPCL RPCCL

11

Intelligent Database Systems Lab

N.Y.U.S.T.

I. M.Conclusion

RPCCL has novelly circumvented the difficult selection of the deleaning rate.

12

Intelligent Database Systems Lab

N.Y.U.S.T.

I. M.

Personal Opinions

Advantage RPCCL can automatically select the correct cluster n

umber. The novel technique can circumvent the selection of t

he delearning rate.

Drawback limitation : k >= k*

Application clustering…

13

Intelligent Database Systems Lab

N.Y.U.S.T.

I. M.K-means example

1. Given : {2,4,10,12,3,20,30,11,25} k=2

2. Randomly assign means : m1=3 ; m2=4

k1={2,3} , k2={4,10,12,20,30,11,25} ,m1=2.5 , m2=16

k1={2,3,4} , k2={10,12,20,30,11,25} , m1=3 , m2=18

k1={2,3,4,10} , k2={12,20,30,11,25} , m1=4.75 , m2=19.6

k1={2,3,4,10,11,12} , k2={20,30,25} , m1=7 , m2=25

…..

14

Intelligent Database Systems Lab

N.Y.U.S.T.

I. M.Dead-unit problem

1. Given : {2,4,10,12,3,20,30,11,25} , k=3

2. Randomly assign means : m1=30 ; m2=25 ; m3=10

Dead-unit

Heuristic Frequency Sensitive Competitive Learning (FSCL) algorithm