Elsevier

Pattern Recognition

Volume 27, Issue 11, November 1994, Pages 1567-1573
Pattern Recognition

Optimality tests for the fuzzy c-means algorithm

https://doi.org/10.1016/0031-3203(94)90134-1Get rights and content

Abstract

Because of lack of convexity of the fuzzy c-means (FCM) objective functionals, the FCM algorithm may converge to a local minimizer or a saddle point. In this paper, we present a new scheme for testing the optimality of the fixed points of the FCM algorithm, one which needs much less computation than Kim et al.'s scheme (Pattern Recognition21, 651–663, 1988). We also point out that Kim et al.'s scheme is partly incorrect, and propose a repair of it. Numerical experiments are presented for several sets of data; they compare Kim et al.'s scheme, the repaired scheme, and our scheme.

References (7)

There are more references available in the full text version of this article.

Cited by (37)

  • Improving Wang-Mendel method performance in fuzzy rules generation using the fuzzy C-means clustering algorithm

    2015, Neurocomputing
    Citation Excerpt :

    In addition, Eqs. (5) and (6) are used in step 2 of the AFCM algorithm to update the degree of affinity, which is considered to be equal in the initialization step. A number of previous studies have investigated the convergence properties of FCM clustering [28–30]. The sample-weighted algorithm is not influenced by outliers, while its superiority and effectiveness have been demonstrated [27].

  • An enhanced classification method comprising a genetic algorithm, rough set theory and a modified PBMF-index function

    2012, Applied Soft Computing Journal
    Citation Excerpt :

    Amongst these methods, traditional indices such as the partition coefficient [6,7] and classification entropy coefficient [8,9] are based simply on the membership values of the items within the dataset and are therefore computationally straightforward. However, recent studies have shown that the accuracy of a cluster validity index can be improved by considering not only the dataset itself, but also the matrix U used to partition the data [10–17]. In general, existing clustering methods cluster the dataset in accordance with the norms of the instances rather than the values of the individual attributes of the instances.

  • Sample-weighted clustering methods

    2011, Computers and Mathematics with Applications
View all citing articles on Scopus
View full text