The paper addresses the problem of estimating multi-label noise transition matrices in noisy multi-label learning. Given the challenges of collecting large-scale accurate labels, the authors propose using transition matrices to model multi-label noise. The proposed method leverages label correlations without requiring anchor points or accurate fitting of noisy class posteriors. The occurrence probability of two noisy labels is first estimated to capture label correlations, followed by employing sample selection techniques to extract clean label correlation information. The effectiveness of the proposed estimator is validated through empirical results demonstrating excellent classification performance.
Publication date: 25 Sep 2023
Project Page: Not Provided
Paper: https://arxiv.org/pdf/2309.12706
