DRN: Detection and Removal of Noisy Instances with Self Organizing Map

Abstract

Identification of noisy instances provides an effective solution to improve the predictive performance of machine learning algorithms. The presence of noise in a data set poses two major negative consequences: (i) a decrease in the classification accuracy (ii) an increase in the complexity of the induced model. Therefore, the removal of noisy instances can improve the performance of the induced models. However, noise identification can be especially challenging when learning complex functions which often contain outliers. To detect such noise, we present a novel approach: DRN for detecting instances with noise. In our approach, we ensemble a self-organizing map (SOM) with a classifier. DRN can effectively distinguish between outlier and noisy instances. We evaluate the performance of our proposed algorithm using five different classifiers (viz. J48, Naive Bayes, Support Vector Machine, $k$-Nearest Neighbor, Random Forest) and 10 benchmark data sets from the UCI machine learning repository. Experimental results show that DRN removes noisy instances effectively and achieves better accuracy than the existing state-of-the-art algorithm on various datasets.