Advisor : Dr. Hsu Presenter : Wen-Hsiang Hu Authors : D.S. Guru*; Bapu B. Kiranagi;

15
Intelligent Database Systems Lab 國國國國國國國國 National Yunlin University of Science and T echnology Multivalued type dissimi larity measure and conce pt of mutual dissimilari ty value for clustering symbolic patterns Advisor : Dr. Hsu Presenter : Wen-Hsiang Hu Authors : D.S. Guru*; Bapu B. Kiranagi; Pattern Recognition Society. Published by Elsevier Ltd, 2004, Pages:151 - 156

description

Multivalued type dissimilarity measure and concept of mutual dissimilarity value for clustering symbolic patterns. Advisor : Dr. Hsu Presenter : Wen-Hsiang Hu Authors : D.S. Guru*; Bapu B. Kiranagi;. Pattern Recognition Society. Published by Elsevier Ltd, 2004, Pages:151 - 156. - PowerPoint PPT Presentation

Transcript of Advisor : Dr. Hsu Presenter : Wen-Hsiang Hu Authors : D.S. Guru*; Bapu B. Kiranagi;

Page 1: Advisor    : Dr. Hsu Presenter  : Wen-Hsiang Hu Authors    : D.S. Guru*;        Bapu B. Kiranagi;

Intelligent Database Systems Lab

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

Multivalued type dissimilarity measure and concept of mutual dissimilarity value for clustering symb

olic patternsAdvisor : Dr. Hsu

Presenter : Wen-Hsiang Hu

Authors : D.S. Guru*;

Bapu B. Kiranagi;

Pattern Recognition Society. Published by Elsevier Ltd, 2004, Pages:151 - 156

Page 2: Advisor    : Dr. Hsu Presenter  : Wen-Hsiang Hu Authors    : D.S. Guru*;        Bapu B. Kiranagi;

2Intelligent Database Systems Lab

N.Y.U.S.T.

I. M.

Outline Motivation Objective Introduction Dissimilarity Measure Modified Agglomerative clustering technique by

introducing the concept of MDV Experiments Conclusion Personal Opinion

Page 3: Advisor    : Dr. Hsu Presenter  : Wen-Hsiang Hu Authors    : D.S. Guru*;        Bapu B. Kiranagi;

3Intelligent Database Systems Lab

N.Y.U.S.T.

I. M.

Motivation For methods [2-5], the degree of proximity between two symb

olic patterns is assumed to be crisp and symmetric.

However, the proximity can, in general, be expected either to lie within a certain range or to be an instance of multivalued type in addition to being non-symmetric.

A special instance of crisp, non-symmetric and Boolean.

Page 4: Advisor    : Dr. Hsu Presenter  : Wen-Hsiang Hu Authors    : D.S. Guru*;        Bapu B. Kiranagi;

4Intelligent Database Systems Lab

N.Y.U.S.T.

I. M.

Objective In order to work on such unconventional proximity value.

Page 5: Advisor    : Dr. Hsu Presenter  : Wen-Hsiang Hu Authors    : D.S. Guru*;        Bapu B. Kiranagi;

5Intelligent Database Systems Lab

N.Y.U.S.T.

I. M.

Introduction We present a novel dissimilarity measure to estimate the degre

e of dissimilarity between two symbolic patterns. The proposed measure unlike other methods [2–5] approximat

es degree of dissimilarity by multivalued type data and in addition, it is non-symmetric.

[2]

[5]

[3][4]

Page 6: Advisor    : Dr. Hsu Presenter  : Wen-Hsiang Hu Authors    : D.S. Guru*;        Bapu B. Kiranagi;

6Intelligent Database Systems Lab

N.Y.U.S.T.

I. M.

A novel dissimilarity measure Let Fi and Fj be two symbolic patterns described by n interval

valued features as follows:

For example:

Page 7: Advisor    : Dr. Hsu Presenter  : Wen-Hsiang Hu Authors    : D.S. Guru*;        Bapu B. Kiranagi;

7Intelligent Database Systems Lab

N.Y.U.S.T.

I. M.

A novel dissimilarity measure (cont.)

The degree of dissimilarity of Fi to Fj , with respect to the kth feature (irrespective of overlapping or no overlapping) is characterized by

Fik Fjk

Page 8: Advisor    : Dr. Hsu Presenter  : Wen-Hsiang Hu Authors    : D.S. Guru*;        Bapu B. Kiranagi;

8Intelligent Database Systems Lab

N.Y.U.S.T.

I. M.Modified agglomerative clustering technique by introducing the concept of MDV

The “Mutual dissimilarity value” between two patterns is defined to be the magnitude of the vector, which is the sum of the scalar times of the vectors representing the degree of dissimilarity between the patterns. i.e.

MDV is derived by the use of non-symmetric values Di→j and Dj→i with different weight factors α and β.

If Di→j and Dj→i are one and the same (as in conventional techniques), the weight factors do not convey any meaning.

where α and β are scalars (weights).

Page 9: Advisor    : Dr. Hsu Presenter  : Wen-Hsiang Hu Authors    : D.S. Guru*;        Bapu B. Kiranagi;

9Intelligent Database Systems Lab

N.Y.U.S.T.

I. M.Modified agglomerative clustering technique by introducing the concept of MDV (cont.)

Initially m clusters

Page 10: Advisor    : Dr. Hsu Presenter  : Wen-Hsiang Hu Authors    : D.S. Guru*;        Bapu B. Kiranagi;

10Intelligent Database Systems Lab

N.Y.U.S.T.

I. M.

For sake of simplicity the weight factors α and β are set to 1.

Experiments

[2]

[5]

[3][4]

[5][2][3][4]

Page 11: Advisor    : Dr. Hsu Presenter  : Wen-Hsiang Hu Authors    : D.S. Guru*;        Bapu B. Kiranagi;

11Intelligent Database Systems Lab

N.Y.U.S.T.

I. M.

Experiments (cont.)

Page 12: Advisor    : Dr. Hsu Presenter  : Wen-Hsiang Hu Authors    : D.S. Guru*;        Bapu B. Kiranagi;

12Intelligent Database Systems Lab

N.Y.U.S.T.

I. M.

Experiments (cont.)

Page 13: Advisor    : Dr. Hsu Presenter  : Wen-Hsiang Hu Authors    : D.S. Guru*;        Bapu B. Kiranagi;

13Intelligent Database Systems Lab

N.Y.U.S.T.

I. M.

In this paper, a novel dissimilarity measure and a modified clustering algorithm by introducing the concept of MDV is proposed for clustering symbolic patterns.

The proposed method is experimentally validated for its efficacy and is shown to have high consistency with human perception.

Conclusion

Page 14: Advisor    : Dr. Hsu Presenter  : Wen-Hsiang Hu Authors    : D.S. Guru*;        Bapu B. Kiranagi;

14Intelligent Database Systems Lab

N.Y.U.S.T.

I. M.

Our method bears the following characteristics: It is simple and computationally efficient. It can be employed on quantitative, qualitative and

multivalued qualitative symbolic data types. It is non-parametric. Being based on MDV it works on multivalued type

proximity matrix.

Conclusion (cont.)

Page 15: Advisor    : Dr. Hsu Presenter  : Wen-Hsiang Hu Authors    : D.S. Guru*;        Bapu B. Kiranagi;

15Intelligent Database Systems Lab

N.Y.U.S.T.

I. M.

Drawback Does not explain what is suitable weight factors ( αand

β) for obtaining a better cluster of symbolic patterns

Application Apply dissimilarity to ART

Personal Opinion