Improving class noise detection and classification performance: a new two-filter CNDC model

Abstract

Class noise is an important issue in classification with a lot of potential consequences. It can decrease the overall accuracy and increase the complexity of the induced model. This study investigates ensemble filtering, removing and relabeling noisy instances issues and proposes a new two-filter model for Class Noise Detection and Classification (CNDC). The proposed two-filter CNDC model comprises two major parts, which are noise detection and noise classification. The noise detection part involves ensemble and distance filtering to overcome ensemble issues. In latter part, a Removing-Relabeling (REM-REL) technique is proposed to enhance overall performance of noise classification. To evaluate the performance of the proposed model, several experiments were conducted on six real data sets. The proposed REM-REL technique was found to be successful to classify noisy instances. The final results showed that the proposed model led to a significant performance improvement compared with ensemble filtering.