Fuzzy rough sets, and a granular neural network for unsupervised feature selection

• Published in: Neural networks 2013-12, Vol.48, p.91-108
• Main Authors: Ganivada, Avatharam, Ray, Shubhra Sankar, Pal, Sankar K.
• Format: Article
• Language: English
• Subjects: Algorithmics. Computability. Computer arithmetics; Algorithms; Applied sciences; Arrhythmias, Cardiac - classification; Artificial Intelligence; Atmosphere; Bayes Theorem; Cell Cycle; Computer science; control theory; systems; Connectionism. Neural networks; Data processing. List processing. Character string processing; Databases, Factual - classification; Decision Theory; Electronic Mail - classification; Entropy; Exact sciences and technology; Feature evaluation; Fuzzy Logic; Granular computing; Humans; Memory organisation. Data processing; Microarray Analysis; Neoplasms - classification; Neural Networks (Computer); Plants - classification; Rough fuzzy computing; Rule based layered network; Semiconductors - classification; Software; Support Vector Machine; Terminology as Topic; Theoretical computing; Wavelet Analysis
• ISSN: 0893-6080; 1879-2782
• DOI: 10.1016/j.neunet.2013.07.008

A granular neural network for identifying salient features of data, based on the concepts of fuzzy set and a newly defined fuzzy rough set, is proposed. The formation of the network mainly involves an input vector, initial connection weights and a target value. Each feature of the data is normalized between 0 and 1 and used to develop granulation structures by a user defined α-value. The input vector and the target value of the network are defined using granulation structures, based on the concept of fuzzy sets. The same granulation structures are also presented to a decision system. The decision system helps in extracting the domain knowledge about data in the form of dependency factors, using the notion of new fuzzy rough set. These dependency factors are assigned as the initial connection weights of the proposed network. It is then trained using minimization of a novel feature evaluation index in an unsupervised manner. The effectiveness of the proposed network, in evaluating selected features, is demonstrated on several real-life datasets. The results of FRGNN are found to be statistically more significant than related methods in 28 instances of 40 instances, i.e., 70% of instances, using the paired t-test.