Stacking Model-Based Classifiers for Dealing With Multiple Sets of Noisy Labels

Supervised learning in presence of multiple sets of noisy labels is a challenging task that is receiving increasing interest in the ever-evolving landscape of healthcare analytics. Such an issue arises when multiple annotators are tasked to manually label the same training samples, potentially giving rise to discrepancies in class assignments among the supplied labels with respect to the ground truth. Commonly, the labeling process is entrusted to a small group of domain experts, and different level of experience and subjectivity may result in noisy training labels. To solve the classification task leveraging on the availability of multiple data annotators, we introduce a novel ensemble methodology constructed combining model-based classifiers separately trained on single sets of noisy labels. Eigenvalue Decomposition Discriminant Analysis is employed for the definition of the base learners, and six distinct averaging strategies are proposed to combine them. Two solutions necessitate a priori information, such as the partial knowledge of the ground truth labels or the annotators' level of expertise. Differently, the remaining four approaches are entirely data-driven. A simulation study and an application on real data showcase the improved predictive performance of our proposal, while also demonstrating the ability of automatically inferring annotators' expertise level as a by-product of the learning process.