TUTORIAL:   Recent Trends in Fuzzy Clustering: Emerging Frontiers of Information Granulation

RESP:  Witold Pedrycz, PhD -  Department of Electrical & Computer Engineering - University of Alberta - Canada

 

ABSTRACT

We are witnessing a significant paradigm shift in fuzzy clustering by moving from plain data-driven pursuits to knowledge-based developments of structures in data and clustering aimed at the discovery of information granules of higher type or higher order. Fuzzy clustering is regarded as a fundamental conceptual and algorithmic framework of designing information granules and in this way the paradigm shift offers new and attractive opportunities.

We first elaborate on a variety of the facets of domain knowledge, which is used to augment and navigate clustering mechanisms. The underlying taxonomy of the generic categories includes: (a) partial labeling, (b) proximity–based navigation hints, (c) viewpoints, and (d) context-based guidance of clustering.

We present a hierarchical way of forming fuzzy clusters and investigate two design strategies as well as demonstrate how the clusters resulting at the higher conceptual level can be represented as fuzzy sets of type-2. In the passive approach, the clusters represented in terms of fuzzy sets of type-2 are constructed on a basis of clustering results available at the lower level. In contrast, in the active approach the clusters being built at the higher conceptual level adjust the clusters at the lower level meaning that there is a feedback loop of structural character between the two layers of the clustering hierarchy.

Discussed is also a way of forming granular rather numeric prototypes and a technique in which such prototypes give rise to type-2 fuzzy sets describing clusters.

The presentation addresses the direct usage of information granules in the design of neural networks, in particular architectures falling under the rubric of Radial Basis Function neurocomputing.