130 likes | 301 Views
A Similarity-Based Robust Clustering Method. Author : Miin-ShenYang and Kuo-Lung Wu Reporter : Tze Ho-Lin 2006/2/8. PAMI, 2004. Outline. Motivation Objectives Methodology Evaluation Conclusion Personal Comments Appendix. Motivation.
E N D
A Similarity-Based Robust Clustering Method Author : Miin-ShenYang and Kuo-Lung Wu Reporter : Tze Ho-Lin 2006/2/8 PAMI, 2004
Outline • Motivation • Objectives • Methodology • Evaluation • Conclusion • Personal Comments • Appendix
Motivation • Most clustering methods are less to include the property of robustness.
Objectives • Construct a robust clustering method that • Robust to the initialization (cluster number and initial guesses) • Robust to cluster volumes (ability to detect different volumes of clusters) • Robust to noise and outliers
γ=1 5 iteration γ=10 converge Methodology
Evaluation Data set FCM PCM SCM with single-link method SCM with Ward’s method SCM convergence state
Evaluation For all n data points in s-dimensional space • CCA • a good estimate of γalways falls in the interval [5,20] • SCA • AHC • PCM & FCM
Conclusion • CCA is used to estimate parameter γ. • SCA is used to self-organize the data • AHC is used to obtain the optimal cluster number c* and identify these c* clusters. • The robustness to different cluster shapes should be another robust clustering characteristic that will be a further research topic.
Personal Comments • Application • Low-dimensional data space clustering • Advantage • SCM can achieve robust clustering results • Disadvantage • Compared with other clustering method, SCM requires more computational time.
Appendix: The Robust properties to noise and outliers (20) (21) φfunction of our estimate
Correlation Comparison Algorithm (CCA) γ=5 γ=10 (7)
5 iteration converge Similarity Clustering Algorithm (SCA) (10) (11) (5)
Fig 5. The identified clusters Fig 4. The Hierarchical Clustering tree Agglomerative Hierarchical Clustering (AHC)