PairwiseLearning

Thefollowing lemma provides moment bounds for a summation of weakly dependent and mean-zero random functions withbounded increments underachange ofanysinglecoordinate [1,10]. The stated bound then follows by combining the above two inequalities together. Note A(S0) is independent ofS and can be considered as a fixed model if we only consider the randomness induced fromS. In this section, we present the proof related to stability and generalization for pairwise learning with convex and smooth loss functions. For anyi [n], define Si as (3.3).