Affiliations: School of Electrical and Computer Engineering, Shiraz University, Shiraz, Iran
Correspondence:
[*]
Corresponding author: Homeira Shahparast, School of Electrical and Computer Engineering, Shiraz University, Shiraz, Iran. E-mail:[email protected]
Abstract: Extracting comprehensive rules from high-dimensional data
is a serious challenge in designing fuzzy classifiers. Among several methods
for generating rules from data, mostly often work efficiently for low
dimensions. Indeed, when dimensions go up, the number of generated rules
becomes unmanageable. In this paper, a feasible approach for extracting
rules from high-dimensional data (FERHD) is proposed. Unlike top-down
methods which generate some general fuzzy rules and then try to make them
specific, our method works in a bottom-up manner. It first generates all
manageable specific rules and then tries to generalize them. In this regard,
after partitioning the problem space into some fuzzy grids, FERHD generates
rules for these partitions if there are at least one training pattern in
their decision subspace. Thus, FERHD is scalable since it generates at most
m rules for a dataset of size m. Also, it decides on suitable number of fuzzy
sets to be used for attributes via the generalization process which in turn
produces a small-size rule base. To justify the scalability of FERHD on
high-dimensional datasets, it is used to extract rules from some benchmark
datasets. In comparing with some related methods, the accuracy and
interpretability of the designed classifiers are acceptable.
