Data preprocessing methods for Sparse Auto-encoder based fuzzy rule classifier
Rahul K. Sevakula, Abhi Shah, Nishchal Kumar Verma · 2015
Sparse Auto-encoders (SA) have lately become very popular for their ability to represent data in a compact form. Due to this property, SAs have been earlier used for fuzzy rule reduction in control, regression and classification problems. The learning capability of SAs depend on the quality of input data, and most often some kind of pre-processing may become a mandatory requisite for learning. This paper proposes data preprocessing methods to improve the performance of SAs during fuzzy rule reduction. The proposed approach enables the SA based fuzzy rule classifier to work on both real, as well as categorical attribute type data sets, and also with improved performance. Proposed methods were tested and compared w.r.t. performance and rule size over 7 data sets. The experimentation provided satisfactory results, and in some cases the proposed model gave improvements in classification accuracy by upto 3 percent and reduction in rule base by over 40 times, as compared to traditional fuzzy rule classifiers.