- Stability-based Generalization Evaluation for Mixtures of Pointwise and Pairwise Studying(arXiv)
Writer : Jiahuan Wang, Jun Chen, Hong Chen, Bin Gu, Weifu Li, Xin Tang
Summary : Not too long ago, some combination algorithms of pointwise and pairwise studying (PPL) have been formulated by using the hybrid error metric of “pointwise loss + pairwise loss” and have proven empirical effectiveness on characteristic choice, rating and advice duties. Nonetheless, to the most effective of our data, the training concept basis of PPL has not been touched within the current works. On this paper, we attempt to fill this theoretical hole by investigating the generalization properties of PPL. After extending the definitions of algorithmic stability to the PPL setting, we set up the high-probability generalization bounds for uniformly secure PPL algorithms. Furthermore, express convergence charges of stochastic gradient descent (SGD) and regularized threat minimization (RRM) for PPL are said by creating the steadiness evaluation strategy of pairwise studying. As well as, the refined generalization bounds of PPL are obtained by changing uniform stability with on-average stability.
