Leveraging local and global relationships for corrupted label detection

Abstract

The performance of the Machine learning and Deep learning models heavily depends on the quality and quantity of the training data. However, real-world datasets often contain a considerable percentage of noisy labels, ranging from 8.0% to 38.5%. This could significantly reduce model accuracy. To address the problem of corrupted labels, we propose Cola, a novel data-centric approach that leverages both local neighborhood similarities and global relationships across the entire dataset to detect corrupted labels. The main idea of our approach is that similar instances tend to share the same label, and the relationship between clean data can be learned and utilized to distinguish the correct and corrupted labels. Our experiments on four well-established datasets of image and text demonstrate that Cola consistently outperforms state-of-the-art approaches, achieving improvements of 8% to 21% in F1-score for identifying corrupted labels across various noise types and rates. For visual data, Cola achieves improvements of up to 80% in F1-score, while for textual data, the average improvement reaches about 17% with a maximum of 91%. Furthermore, Cola is significantly more effective and efficient in detecting corrupted labels than advanced large language models, such as Llama3, with improvements of up to 112% in Precision and a 300X reduction in execution time.