Nonlinear Multiple-Annotators Classifier Based on Expectation-Maximization Under Noisy Labels
W. Cárdenas-Bedoya, Julián Gil-González, David Augusto Cardenas Peña · 2025
In many real-world scenarios, obtaining reliable ground truth labels is expensive and often unfeasible. As a result, machine learning systems frequently rely on multiple annotators whose levels of expertise can vary significantly. This variability introduces label noise, which poses a major challenge for traditional machine learning models that assume clean and consistent annotations from a single expert, often resulting in degraded performance under such conditions. To handle this issue, we propose an Expectation-Maximization-based Nonlinear Multiple Annotators (termed ENMA) classification methodology that combines a flexible neural architecture with the Expectation-Maximization (EM) algorithm. The model employs a backbone encoder mapping inputs into a latent space, enabling the estimation of sample-dependent annotator reliability and robust classification simultaneously. Unlike previous approaches that assume linear separability or uniform annotator expertise, our method adapts to nonlinear relationships and varying annotator behavior across samples. Experimental results on synthetic datasets and the real-world Ionosphere dataset show that the proposed model outperforms both linear and nonlinear baselines in terms of predictive performance and robustness, particularly in imbalanced and high-dimensional settings.