SOMbased Data Visualization Methods Author Juha Vesanto Advisor

SOM-based Data Visualization Methods Author: Juha Vesanto Advisor: Dr. Hsu Graduate: Zen. John Huang 2002/1/24 IDSL seminar 2002/01/24 IDS Lab seminar

Outline Motivation Objective Introduction Methods SOM visualization Conclusions 2002/1/24 IDS Lab seminar 2

Motivation Data mining Complexity or amount of data is prohibitively large for human observation alone An interactive process 2002/1/24 IDS Lab seminar 3

Objective To give an idea What kind of information can be acquired from different presentations How the SOM can best be utilized in exploratory data visualization 2002/1/24 IDS Lab seminar 4

Introduction SOM(self-organizing map) A neural network algorithm based on unsupervised learning A valuable tool in data mining and KDD Applications in • • • 2002/1/24 Full-text Financial data analysis Pattern recognition Image analysis Process monitoring Fault diagnosis IDS Lab seminar 5

SOM Grid 1 - or 2 -dimension Hexagonal or rectangular 2002/1/24 IDS Lab seminar 6

SOM (Cont’d) mk : = mk + α(t) hck(t) (x-mk) α(t) is learning rate hck(t) is a neighborhood kernel centered on the winner unit c 2002/1/24 IDS Lab seminar 7

Some Vector quantization Algorithms 2002/1/24 IDS Lab seminar 8

Some Vector Projection Algorithms 2002/1/24 IDS Lab seminar 9

Different Between SOM and Other Methods Be not serial combination SOM has a regularly shaped projection grid 2002/1/24 IDS Lab seminar 10

Disadvantages of the Rigid Grid The grid guides the vector quantization process The axes of the map grid rarely have any clear interpretation The projection implemented by the SOM alone if very crude 2002/1/24 IDS Lab seminar 11

Projecting Prototype Vectors to a Low Dimension 2002/1/24 IDS Lab seminar 12

Cluster Structure of the SOM 2002/1/24 IDS Lab seminar 13

Component Planes and Histograms 2002/1/24 IDS Lab seminar 14

Cluster Properties |mik – mjk| / ||mi – mj|| k: component i, j: two neighboring map units mi: a prototype vector 2002/1/24 IDS Lab seminar 15

Contribution of News Paper 2002/1/24 IDS Lab seminar 16

Component and Reorganized Planes 2002/1/24 IDS Lab seminar 17

Scatter Plot and Color Map 2002/1/24 IDS Lab seminar 18

Different Ways to Visualize Data Histograms 2002/1/24 IDS Lab seminar 19

All and Scandinavian Mills 2002/1/24 IDS Lab seminar 20

Response Surfaces 2002/1/24 IDS Lab seminar 21

Quantization Error Plots 2002/1/24 IDS Lab seminar 22

CCA-like Projection Algorithm 2002/1/24 IDS Lab seminar 23

Conclusions Bringing the many visualization methods for SOM together Using the software package for Matlab 5 computing environment by Mathworks 2002/1/24 IDS Lab seminar 24

Future Work Some areas may be discarded as outliers Postprocessing 2002/1/24 IDS Lab seminar 25