Weighted Cluster Ensembles: Methods and analysis

1. Weighted Cluster Ensembles:Methods and analysis Presenter：Chien-Hsing Chen Author: Carlotta Domeniconi Muna Al-Razgan 2009.TKDD.40..

2. Outline • Motivation • Objective • Overall of clustering ensemble • Method • Experiments • Conclusion • Comment

3. Motivation • High-dimensional • A dimension (feature) is highly relevant to a cluster, but is irrelevant to another cluster. • Common global dimensionality reduction techniques are unable to capture such local structure of the data. • it instead of • using an equal weight for all w1, w2, …, wD. • using an equal weight for a wi among all clusters, where i=1, …, D, • Clustering ensemble • An ensemble bag includes: K-means, SOM, … etc • Alternative bag is: 3-means, 5-means, 7-means • How can a technique combine the two respects? w=(0.9, 0.8, 0.1)t c1={sport} attribute name homerun baseball shopping w=(0.1, 0.2, 0.9)t c2={auction} w1,i ≠ w2,i

4. Objective • High-dimensional • provide a first attempt to capture local structure of the data. • LAC-h approach • Clustering ensemble • LAC-1, LAC-3, LAC-29, … • Combine the two respects • WSPA approach • WBPA approach • WSBPA approach w1,i ≠ w2,i

5. clustering partition Overall work • 1. A new clustering • approach is discussed • handle high-D • Clustering ensemble • 2. Three ensemble • techniques are introduced • consensus function s ( ) > s ( ) 0.25 0.15 0.95 3. Graph cut 0.01 0.13 0.20 0.91 0.20

6. LAC (locally adaptive clustering) • Clustering ensemble • distance of a attribute i within a cluster j c1 w=(0.9, 0.5, 0.1)t ? |nc1| = 4 c2 w=(0.1, 0.5, 0.9)t |nc2| = 3

7. Overall work • 1. A new clustering • approach is discussed • handle high-D • Clustering ensemble c1 • 2. Three ensemble • techniques are introduced • consensus function s ( ) > s ( ) 0.25 0.15 0.95 3. Graph cut 0.01 0.13 0.20 0.91 0.20

8. WSPA 1/2 0.04 0.94 0.02 0.06 0.90 0.04 s ( ) P =(0.94, 0.04, 0.02)t P =(0.90, 0.06, 0.02)t

9. WSPA 2/2 • Clustering ensemble • Two points have high similarity score if often appearing in the same partitions. • Instance-based Graph cut 0.25 0.15 0.13 0.95 0.01 0.20 0.91 0.20

10. WBPA 1/3 • Problem definition … 0.04 0.02 0.91 0.94 0.03 0.06 P =(0.94, 0.04, 0.02)t P =(0.03, 0.91, 0.06)t • and are never clustered together ≡ 0 Graph • the groups to which and belong share the same instances

11. WBPA 2/3 Graph The Graph is connect between a cluster and an instance instead of that among data

12. WBPA 3/3 0.64 0.94

13. WSBPA 0.94 0.94 0.93 0.91 0.86 0.86 0.85 0.89 0.93 0.94 0.64

14. WSBPA 0.04 0.04 0.03 0.01 0.86 0.86 0.85 0.89 0.01 0.89 0.94 0.64

15. Experiment

16. Experiment

17. Experiment

18. Experiment

19. Experiment

20. w1,i ≠ w2,i Experiment

21. Experiment

22. Conclusion • High-dimensional • LAC-h approach • Clustering ensemble • LAC-1, LAC-3, LAC-29, … • Combine the two respects • WSPA approach • WBPA approach • WSBPA approach

23. Comment • Advantage • Consensus function • Drawback • Application • Ensemble clustering on SOM