Transcript Neural Network Homework Report: Clustering of the Self

Neural Network Homework Report:
Clustering of the Self-Organizing Map
IEEE TRANSACTIONS ON NEURAL NETWORKS, VOL.11, NO.3, MAY 2000
Professor：Hahn-Ming Lee
Student : Hsin-Chung Chen
M9315928
OUTLINE





INTRODUCTION
CLUSTERING
SOM CLUSTERING
EXPERIMENTS
CONCLUSION
INTRODUCTION

DATA mining processes







problem definition
data acquisition.
data preprocessing and survey
data modeling
evaluation.
knowledge deployment.
Self-organization map feature:




Dimensionality reduction of unsupervised learning
Can applied in deal huge amounts of sample
The original data set is represented using a smaller set of prototype
vectors
not to find an optimal clustering but to get good
CLUSTERING

two main ways approaches

hierarchical approaches

agglomerative algorithm：


divisive algorithm：


top-down strategies to build a hierarchical clustering tree
partitive approaches


bottom-up strategies to build a hierarchical clustering tree
k-means
optimal clustering is a partitioning


minimizes distances within
maximizes distances between clusters
CLUSTERING(cont.)
SOM CLUSTERING

SOM training

first to find the best matching unit (BMU)

the prototype vectors are updated.

The SOM algorithm characteristic：



applicable to large data sets.
The computational complexity scales linearly
with the number of data samples
it does not require huge amounts of memory
that basically just the prototype vectors and the
current training vector .
EXPERIMENTS


Tools: SOM_ToolBox 2.0 :
Data set: clown.dat

Data set (“clown.data”) consisted of 2220 2-D samples.






cluster with three subclusters (right eye)
spherical cluster (left eye)
elliptical cluster (nose)
nonspherical cluster (U-shaped: mouth)
large and sparse cluster (body)
noise .(such as black x)

Methods and Parameters:

Cluster step 1:



Training Parameters of the SOM's
Map size: 19x17
Initial Neighborhood Widths:
Rough Phases σ1(0): 10
Fine-Tuning Phases σ2(0): 2
learning rates:(The learning rate decreased linearly to zero during
the training)
Rough Phases : 0.5
Fine-Tuning Phases 0.05
Cluster step 2:

Method: K-Means
Using 100 Runs
Experimental Results
single linkage dendrogram of 323 SOM Map unit
SOM Map average linkage
dendrogram of 323 SOM Map unit
complete linkage dendrogram of 323
SOM Map unit
Conclusion