Fuzzy Self-Organizing Incremental Neural
Network for Fuzzy Clustering
Tianyue Zhang, Baile Xu, and Furao Shen
(
B
)
National Key Lab oratory for Novel Software Technology,
Department of Computer Science and Technology,
Collab orative Innovation Center of Novel Software Technology and Industrialization,
Nanjing University, Nanjing, China
njucszty@gmail.com, dg1633021@smail.nju.edu.cn, frshen@nju.edu.cn
Abstract. In this paper, a neural network named fuzzy self-organizing
incremental neural network (fuzzy SOINN) is presented for fuzzy clus-
tering with following four characteristics: fuzzy, incremental learning,
top ological representation and resistance to noise. No predefined struc-
tures of clusters is required due to the self-adjusting nodes and edges
which fit the learning data incrementally. A removal of nodes and edges
promises the robustness of the network to the noisy data. Experiments on
artificial and real-world data prove the validity of the clustering method.
Keywords: Fuzzy clustering
· Incremental or online learning · Topolog-
ical representation
· Self-organizing incremental neural network (SOINN)
1 Introduction
Clustering is assigning objects to clusters which have higher similarity in the
same cluster and dissimilarity between the different clusters with an exten-
sive application in diverse research fields [1]. However, the issue of bridges
(or overlaps) between clusters is often encountered in the procedure of clus-
tering according to Nagy [2] which is hard to be solved by hard clustering. The
idea that assigning patterns with grades of membership rather than clustering
them into disjoint clusters, which was introduced and termed fuzzy set by Zadeh
[3], exploited into clustering and termed fuzzy clustering by Ruspini [4], is more
appropriate for dealing with the issue. When data clusters have vague bound-
aries and the fuzzy characteristic of data structures needs to be reserved, fuzzy
clustering methods work well. Applications of fuzzy clustering to problems of
clustering, feature selection, and classifier design have been reported in biology,
medicine, psychology, economics, and many other disciplines.
Previous researches on fuzzy clustering leads to the most popular classical
algorithm: fuzzy C-means (FCM, also termed fuzzy K-means) [5] with frequent
applications in clustering for its simplicity and computational efficiency. Though
classical batch algorithm, FCM has developed its on-line version. For very large
data sets, first online fuzzy c-means is presented in [6] and two online fuzzy
c
Springer International Publishing AG 2017
D. Liu et al. (Eds.): ICONIP 2017, Part I, LNCS 10634, pp. 24–32, 2017.
https://doi.org/10.1007/978-3-319-70087-8
_3