Researchers have developed a novel machine learning approach to handle expert labels that exhibit disagreement or ambiguity, particularly when experts provide interval targets instead of exact values. This method preserves individual expert intervals and models them using a mixture of Beta distributions with a Cramér-distance objective. It also decomposes predictive uncertainty into within-component, between-component, and model uncertainty, aligning these components with their corresponding label-side sources through a process called decomposition matching. In tests on sea-ice concentration data, the model achieved a 31% reduction in Mean Absolute Error compared to hard labels and surpassed existing aggregation, interval-distribution, and interval-regression baselines. AI
IMPACT This research offers a more robust way to train models when expert data is inherently uncertain or varied, potentially improving performance in domains with subjective or imprecise expert labeling.
RANK_REASON Academic paper detailing a new machine learning method. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
