Learning from Personal Preferences

Machine learning practitioners frequently use majority vote to resolve disagreement in multi-annotator datasets. While this approach is natural in settings where a single ground truth label exists for each instance, it hides the presence of disagreement for subjective annotation tasks. In domains such as language modeling, information retrieval, and top-k recommendation, models must avoid suppressing minority views and express when the answer to a query is contentious. We propose personalized error metrics to formalize the requirement of strong performance across a heterogeneous user population. Following this framework, we develop an algorithm for training an ensemble of models, each specialized for a different segment of the population.