Feature-Aware Confident Learning to Improve Cloud Revenue Conversion: Leveraging Feature Dependencies for Label Noise Correction

Authors: Pavan Mullapudi

DOI: https://doi.org/10.5281/zenodo.15054771

Short DOI: https://doi.org/g884bw

Country: USA

Full-text Research PDF File:   View   |   Download

Abstract: Label noise is a prevalent challenge in supervised learning, often degrad- ing model performance. Existing approaches, such as confident learning[4], assume label noise is independent of input features, which can limit their effectiveness in real-world datasets where noise correlates with features. In this paper, we propose Feature-Aware Confident Learning (FACLe), a novel method that models label noise as a function of input features. By dynam- ically estimating a noise transition matrix conditioned on features, FACLe enables the correction of feature-dependent label noise. The method inte- grates unsupervised clustering methods and confident learning to identify noisy samples to learn the feature-conditioned noise patterns. We apply this technique on a sample dataset in the domain of cloud computing. Our experiments demonstrate that FACLe achieves substantial improvements over baseline methods, with an average precision improvement of 15% , equating to an improvement of 10% in revenue conversion.

Keywords:

Paper Id: 232270

Published On: 2023-10-03

Published In: Volume 11, Issue 5, September-October 2023

Share this