Stability-Based Generalization Analysis for Mixtures of Pointwise and Pairwise Learning

• DOI: 10.1609/aaai.v37i8.26205
• Publication Date: 2023-06-27T17:33:57Z
• Authors: Jiahuan Wang; Jun Chen; Hong Chen; Bin Gu; Weifu Li; Xin Tang
• Keywords: Pointwise; Stochastic Gradient Descent; Pointwise convergence; Empirical risk minimization

Recently, some mixture algorithms of pointwise and pairwise learning (PPL) have been formulated by employing the hybrid error metric “pointwise loss + loss” shown empirical effectiveness on feature selection, ranking recommendation tasks. However, to best our knowledge, theory foundation PPL has not touched in existing works. In this paper, we try fill theoretical gap investigating generalization properties PPL. After extending definitions algorithmic stability setting, establish high-probability bounds for uniformly stable algorithms. Moreover, explicit convergence rates stochastic gradient descent (SGD) regularized risk minimization (RRM) are stated developing analysis technique learning. addition, refined obtained replacing uniform with on-average stability.