Probability Estimation for Multi-class Classification Based on Label Ranking

摘要: We consider the problem of probability estimation in setting multi-class classification. While this has already been addressed literature, we tackle it from a novel perspective. Exploiting close connection between and ranking, our idea is to solve former on basis latter, taking advantage recently developed methods for label ranking. More specifically, argue that Plackett-Luce ranking model very natural choice context, especially as can be seen multinomial extension Bradley-Terry model. The latter provides pairwise coupling techniques, which arguably constitute state-of-the-art estimation. explore relationship ranking-based approach estimation, both formally empirically. Using synthetic real-world data, show method does not only enjoy nice theoretical properties, but also competitive terms accuracy efficiency.