Controlling Asymmetric Errors in Neuro-Fuzzy Classification

Aljoscha Klose, Rudolf Kruse
Department of Computer Science (FIN-IWS), University of Magdeburg
Universitätsplatz 2, D-39106 Magdeburg,
Germany Karsten Schulz, Ulrich Thönnessen
FGAN/FOM
Eisenstockstr. 12, 76275 Ettlingen,
Germany

Abstract:

In many practical classification problems the severeness of misclassifications depends on the semantics of true and predicted class in the underlying domain. We present such a problem from machine vision, where additionally the class probabilities are extremely unbalanced. Due to their interpretability neuro-fuzzy classifiers are a popular way to extract rules from example data. However, like most classifier approaches, neuro-fuzzy systems have problems when learning from asymmetric or unbalanced data. We present modifications to the learning algorithm of the NEFCLASS system to cope with these problems, and show experimental results.

Keywords: neuro-fuzzy classifiers, asymmetric errors, unbalanced data, aerial images.

• Introduction
  • Classification of Primitive Objects in SAR Images
  • The NEFCLASS Model
• Extending the NEFCLASS System
  • Modifications to the 1st Learning Phase
  • Modifications to the 2nd Learning Phase
  • Adapting the Pruning Techniques
• Experimental Evaluation
• Discussion
• References
• About this document ...