Learning rate of magnitude-preserving regularization ranking with dependent samples

摘要

The generalization analysis is key to understand the theoretical foundation of learning to rank. However, the previous works for this subject are usually based on independent and identical distributed (i.i.d) samples. In this paper, we go beyond this restriction by investigating the generalization ability of magnitude-preserving regularization ranking (MPRank) with dependent samples. For the MPRank, we establish its upper bound for the excess ranking risk which demonstrates the satisfactory learning rate can be reached for dependent samples.