Title :
An improved differential evolution and its application to determining feature weights in similarity based clustering
Author :
Chun-Ru Dong ; Yeung, Daniel S. ; Xi-Zhao Wang
Author_Institution :
Machine Learning & Cybern. Res. Center, South China Univ. of Technol., Guangzhou, China
Abstract :
Feature weighting, which is considered as an extension of feature selection techniques, has been successfully applied to improve the performance of clustering. Focusing on the clustering based on a similarity matrix, we design an optimization model to minimize the fuzziness of similarity matrix by learning feature weights. The objective of this model is to get a more reasonable result of clustering through minimizing the uncertainty (fuzziness and non-specificity) of similarity matrix. To solving this optimization model effectively, we propose a new searching approach which integrates together multiple evolution strategies of both differential evolution and dynamic differential evolution. The experimental results on several benchmark datasets show that the performance of the proposed method is significantly improved compared to that of gradient-descent-based approach in terms of five selected clustering evaluation indices, i.e., fuzziness of similarity matrix, intra-class similarity, inter-class similarity, ratio of intra-class similarity to inter-class similarity.
Keywords :
feature selection; matrix algebra; optimisation; pattern clustering; dynamic differential evolution; feature selection; feature weights; optimization model; similarity based clustering; similarity matrix; uncertainty minimization; Abstracts; Noise measurement; Differential Evolution; Dynamic Differential Evolution; Feature weights learning; Similarity-based clustering;
Conference_Titel :
Machine Learning and Cybernetics (ICMLC), 2013 International Conference on
Conference_Location :
Tianjin
DOI :
10.1109/ICMLC.2013.6890399