Learning Bregman Distance Functions for Structural Learning to Rank

摘要: We study content-based learning to rank from the perspective of distance functions. Standardly, two key issues rank, feature mappings and score functions, are usually modeled separately, is restricted modeling a linear function such as Mahalanobis distance. However, functions mutually interacted, patterns underlying data probably complicated nonlinear. Thus, general nonlinear family, Bregman suitable for due its strong generalization ability nonlinearity exploring distributions. In this paper, we structural problem, devise build ranking model based on SVM. To improve robustness outliers, develop robust framework model. The proposed Robust Structural Learning Rank ( RSBLR ) unified rank. experiments real-world datasets show superiority method state-of-the-art literature, well noisily labeled outliers.