Analytical and Numerical Evaluation of the Suppressed Fuzzy C-Means Algorithm
MDAI '08 Sabadell Proceedings of the 5th International Conference on Modeling Decisions for Artificial Intelligence
A generalized approach to the suppressed fuzzy c-means algorithm
MDAI'10 Proceedings of the 7th international conference on Modeling decisions for artificial intelligence
Fuzzy-possibilistic product partition: a novel robust approach to c-means clustering
MDAI'11 Proceedings of the 8th international conference on Modeling decisions for artificial intelligence
Weighted Fuzzy-Possibilistic C-Means Over Large Data Sets
International Journal of Data Warehousing and Mining
Expert Systems with Applications: An International Journal
Hi-index | 0.00 |
The possibility based clustering algorithm PCM was first proposed by Krishnapuram and Keller to overcome the noise sensitivity of algorithm FCM (Fuzzy C-Means). However, PCM still suffers from the following weaknesses: (1) the clustering results are strongly dependent on parameter selection and/or initialization; (2) the clustering accuracy is often deteriorated due to its coincident clustering problem; (3) outliers can not be well labeled, which will weaken its clustering performances in real applications. In this study, in order to effectively avoid the above weaknesses, a novel enhanced PCM version (EPCM) is presented. Here, at first a novel strategy of flexible hyperspheric dichotomy is proposed which may partition a dataset into two parts: the main cluster and auxiliary cluster, and is then utilized to construct the objective function of EPCM with some novel constraints. Finally, EPCM is realized by using an alternative optimization approach. The main advantage of EPCM lies in the fact that it can not only avoid the coincident cluster problem by using the novel constraint in its objective function, but also has less noise sensitivity and higher clustering accuracy due to the introduction of the strategy of flexible hyperspheric dichotomy. Our experimental results about simulated and real datasets confirm the above conclusions.