ANCES: A novel method to repair attribute noise in classification problems. (January 2022)

• Record Type: Journal Article
• Authors: Sáez, José A.; Corchado, Emilio
• Is Part Of: Pattern recognition. Volume 121(2022)
• Publication Date: 2022-01
• Subjects: Attribute noise -- Noise correction -- Noise filtering -- Noisy data -- Classification; Pattern perception -- Periodicals; Perception des structures -- Périodiques; Patroonherkenning; 006.4
• DOI: 10.1016/j.patcog.2021.108198
• Languages: English
• ISSNs: 0031-3203

Abstract

Highlights: The removal of samples is inadvisable dealing with data with attribute noise. Correcting attribute values with errors is the best approach in this scenario. This paper proposes a novel attribute noise correction method, called ANCES. ANCES performs an iterative correction of attribute values using error scores. ANCES outperforms other existing noise filtering techniques in the literature. Graphical abstract: Abstract: Noise negatively affects the complexity and performance of models built in classification problems. The most common approach to mitigate its consequences is the usage of preprocessing techniques, known as noise filters, which are designed to remove noisy samples from the training data. Nevertheless, they are specifically oriented to deal with errors affecting class labels. Their employment may not always result in an improvement when noise affects attribute values. In these cases, correcting the errors is an interesting alternative to traditional noise filtering that has not been enough studied so far in the specialized literature. This research proposes an attribute noise correction method with the final aim of increasing the performance of the classification algorithms used later. The identification of noisy data is based on an error score assigned to each one of the attribute values in the dataset, which are then passed through an optimization process to correct their potential noise. The validity of the proposed method is studied in an exhaustive experimental study, in which it is compared to several well-known preprocessing methods to deal with noisy datasets. The results obtained show the suitability of attribute noise correction with respect to the other alternatives when data suffer from attribute noise.