Affiliations: [a] LRIT, Associated Unit to CNRST-URAC no. 29, Faculty of Science, Rabat IT Center, Mohammed V University in Rabat, Morocco
| [b] LGS, National School of Applied Sciences (ENSA), Ibn Tofail University, Kenitra, Morocco
Correspondence:
[*]
Corresponding author. Btissam Zerhari, LRIT, Associated Unit to CNRST-URAC no. 29, Faculty of Science, Rabat IT Center, Mohammed V University in Rabat, Morocco. E-mail: [email protected].
Abstract: Attribute and class noises are the two important sources of Corruptions (noise) contained in real-world datasets which may deteriorate data interpretation and accuracy. Class noise has potentially serious negative impacts compared to attribute noise, however, the existing major class noise detection methods are not able to address this problem efficiently. To overcome issues related to detection and the elimination of class noise, we suggest a new noise filtering approach able to identify and remove class noise, called Multi-Iterative Partitioning Class Noise Filter (MIPCNF). Since there is no single filter that consistently outperforms its counterparts in all database types and in different levels of noise, our approach relies on an algorithm in which several rounds of class noise detection are performed on different partitions of the data using several classifiers. Therefore, we use different filtering strategies: iterative noise filter, partitioning filter and ensemble-based filter. The experimental results, on 14 real-world datasets, and statistical analysis, show that our method is not only overcoming the higher noise but also over-performing latest class noise detection and elimination strategies in different levels of noise.
Keywords: Class noise, Noise Detection, Noise Elimination, Partitioning Filter, Large Data
