Robust Plackett-Luce model for \(k\)-ary crowdsourced preferences (Q1640559)

This paper presents ROPAL, a robust Plackett-Luce model to aggregate k-ary preferences. A denoising vector based on Kendall tau distance is employed to deal with crowdsourced preferences. An online Bayesian inference makes the method scalable for large-scale preferences. Besides distinguishing noisy workers from experts, the model categorizes noisy workers into amateur, spammer and malicious workers. Robustness is demonstrated in experiments conducted on large-scale simulated datasets and on the real-world datasets ``PeerGrading'' and ``BabyFace''.