Learning from Multiple Annotators: An Ensemble Model-Based Classification Approach

Supervised learning with multiple sets of noisy labels\r\npresents a complex challenge, arising when several annotators are\r\nrequired to manually label the same training samples, potentially resulting\r\nin inconsistencies in class assignments compared to the ground truth.\r\nTo efficiently learn a classifier in this context, an ensemble approach is\r\ndeveloped by leveraging model-based discriminant analysis trained individually\r\non distinct sets of noisy labels. Several strategies are proposed\r\nto combine the base learners, extending solutions proposed in the literature\r\nfor the binary classification setting to the multi-class framework.\r\nAn application involving the identification of gastrointestinal lesions from\r\ncolonoscopic videos, revised by seven clinicians, demonstrates the applicability\r\nof our proposal.