Title: Stability and generalization of stochastic gradient descent for pairwise learning
Authors: Yiming Ying - State University of New York at Albany (United States) [presenting]
Abstract: Pairwise learning refers to a learning task which involves a loss function depending on pairs of instances. Most notable examples of pairwise learning include bipartite ranking, metric learning, AUC maximization and minimum error entropy (MEE) principle. We establish the stability results for SGD algorithms for pairwise learning in both convex, strongly-convex and non-convex settings. As a consequence, we derive their generalization error bounds. Finally, we describe our stability results by illustrating some specific examples of pairwise learning such as AUC maximization, metric learning and MEE. The motivation comes from a previous recent work and the results we obtain complement it in the setting of pointwise learning.