From Social to Individuals: a Parsimonious Path of Multi-level Models for Crowdsourced Preference Aggregation

摘要: In crowdsourced preference aggregation, it is often assumed that all the annotators are subject to a common or social utility function which generates their comparison behaviors in experiments. However, reality variations due multi-criteria, abnormal, mixture of such behaviors. this paper, we propose parsimonious mixed-effects model, takes into account both fixed effect majority follows linear and random some might deviate from significantly exhibit strongly personalized preferences. The key algorithm paper establishes dynamic path individual variations, with different levels sparsity on personalization. based Linearized Bregman Iterations, leads easy parallel implementations meet need large-scale data analysis. unified framework, three kinds models presented, including basic model L2 loss, Bradley-Terry Thurstone-Mosteller model. validity these multi-level supported by experiments simulated real-world datasets, shows improvements interpretability predictive precision compared traditional HodgeRank.