Robust Clustering Algorithms and Potential Applications

Several novel and robust learning algorithms, with
the aim to overcome the drawbacks of traditional
clustering algorithms, are developed for data
clustering and its applications. The effectiveness
and superiority of the proposed methods are
supported by experimental results.
1) Te proposed RDA exhibits several robust
clustering characteristics: robust to the
initialization; robust to cluster volumes; and
robust to noise and outliers.
2) The proposed IFCSS algorithm achieves two robust
clustering characteristics: the robustness
against noisy points is obtained by the maximization
of mutual information; and the optimal cluster
number is auto-determined by the VC-bound induced
cluster validity.
3) The KDA is developed to discover some complicated
(e.g., linearly nonseparable) data structures which
can not be revealed by traditional clustering
methods in the standard Euclidean space.
4) Finally, robust clustering methods have been
developed for image segmentation and pattern
classification. The proposed ASDA can perform
unsupervised clustering for robust image
segmentation. The KPCM is developed to generate
weights used for SVM training.

Autorentext

XuLei YANG obtained the PhD degree from EEE School, NTU in 2005. His current research interests include pattern recognition, image processing, and machine vision. He has published more than 20 papers in scientific book chapters, journals and conference proceedings.

Klappentext

Several novel and robust learning algorithms, with the aim to overcome the drawbacks of traditional clustering algorithms, are developed for data clustering and its applications. The effectiveness and superiority of the proposed methods are supported by experimental results. 1) Te proposed RDA exhibits several robust clustering characteristics: robust to the initialization; robust to cluster volumes; and robust to noise and outliers.2) The proposed IFCSS algorithm achieves two robust clustering characteristics: the robustness against noisy points is obtained by the maximization of mutual information; and the optimal cluster number is auto-determined by the VC-bound induced cluster validity.3) The KDA is developed to discover some complicated (e.g., linearly nonseparable) data structures which can not be revealed by traditional clustering methods in the standard Euclidean space.4) Finally, robust clustering methods have been developed for image segmentation and pattern classification. The proposed ASDA can perform unsupervised clustering for robust image segmentation. The KPCM is developed to generate weights used for SVM training.