Integrated label correction for e-commerce data: Boosting accuracy with baseline attention and enhanced Bayesian updating

Abstract

The rapid growth of the e-commerce industry has led to an explosion of product data, which e-commerce platforms increasingly leverage to enhance operational decision-making. However, the presence of label noise in product data poses a significant challenge, as mislabeled data can degrade decision-making performance, negatively impacting both platform efficiency and user experience. To tackle this challenge, this paper proposes an integrated label correction (ILC) method, featuring two key innovations: a baseline attention (BA) mechanism for noise detection and an enhanced Bayesian updating (EBU) strategy for label prediction. The proposed BA mechanism utilizes a dynamic attention scaling approach with learnable parameters to measure the similarity between samples and their labeled classes. The EBU strategy integrates sample features and noisy observations into a unified probabilistic framework, jointly optimizing the classifier and the label correction process. A novel adaptive noise transition learning approach is further introduced to refine the noise transition matrix. Extensive experiments on a real-world JD text dataset, a benchmark image dataset, and an automotive user review dataset validate the effectiveness of the proposed ILC method. It consistently outperforms state-of-the-art approaches on average, achieving a 2.78% improvement on the JD dataset, and gains of 0.80% on the image dataset and 1.73% on the review dataset across different noise types. These improvements demonstrate that the ILC method not only improves classification performance but also offers scalable solutions to reduce mislabeling errors in real-world e-commerce platforms, thereby enhancing recommendation systems, inventory management, and user satisfaction.