Skip to Main Content
Noise is common in any real-world dataset and may adversely affect the accuracy, construction time and complexity of the classifiers. Fuzzy Rule-Based Classification Systems build classifiers widely recognized for their interpretability but also for their good performance and robustness when dealing with imperfect data. This paper analyzes the behavior of such systems with respect to classic crisp systems in the presence of noise. We study the performance of these systems and their robustness in terms of the performance degradation and the size of the classifiers when the noise level increases in training data. We also study their performance when noise is only present in test data the result of a training data polishing and their synergy with several noise filters. The results obtained show that Fuzzy Rule-Based Classification Systems obtain better performance results than crisp systems when they are trained with noisy data. Their robustness results are particularly notable when noise is more disruptive. Their evaluation on noisy test data and their combined use with noise filters are also shown to be appropriate.