An investigation into the effects of label noise on Dynamic Selection algorithms

Walmsley, Felipe N.; Cavalcanti, George D. C.; Sabourin, Robert; Cruz, Rafael M. O.

doi:10.1016/j.inffus.2021.10.015

Cited by 5 publications

(2 citation statements)

References 40 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…They consider that instances with the kDN greater than 0.8 (using k = 5) should not be considered since the 80% of their neighbors are not from their same class and, therefore, are noisy instances that can potentially harm the system. In [31], they carried out an analysis of how noise affects the performance of Dynamic Selection algorithms. They used the kDN to avoid noisy instances by giving more probability to be selected to the instances having a lower value of the kDN .…”

Section: State-of-the-artmentioning

confidence: 99%

“…For example, it has been successfully applied as a noise filter in an online scheme [23]. In the same way, in [31], the authors employ the kDN by filtering the instances with a high value since they can be safely discarded because they are either noisy instances or outliers. Other works utilize the kDN as a guide for informed sampling to improve the performance of classification models [28].…”

Section: Introductionmentioning

confidence: 99%

See 1 more Smart Citation

Dynamic Disagreeing Neighbors: A complexity measure for instance selection

Aceña,

Lancho,

de Diego

et al. 2024

Preprint

View full text Add to dashboard Cite

Within the scope of Machine Learning (ML), Instance Selection (IS) is a sampling process that consists of filtering noise and removing redundant data. In classification problems, IS implies a compromise between maximizing performance and reducing the sample size used for training. Complexity measures provide relevant information about the difficulty of classifying instances, which makes them appropriate for IS since they can capture, for example, noisy or borderline points. This paper introduces the complexity measure DDN, defined at three levels: instance, class, and dataset. The Dynamic Disagreeing Neighbors (DDN) of an instance is defined as the percentage of its nearest neighbors that belong to other classes. The DDN is based on the Nearest Centroid Neighbors (NCN) neighborhood computation, which is dynamically adjusted to the data distribution. In addition, the distance of each neighbor is taken into account so that those farther away are less influential and those closer are more influential for the instance complexity. The validity of the proposal is evaluated through a series of experiments where it is compared with the widely known k-Disagreeing Neighbors (kDN) in terms of stability, correlation with classification error, and performance in IS. The DDN has shown competitive, stable, and robust results throughout the experiments, generally improving on those obtained with the alternative.

show abstract

Section: State-of-the-artmentioning

confidence: 99%