学习算法的稳定性与泛化(Ⅱ):有界差分稳定框架

Stability and generalization of the learning algorithm(Ⅱ):the framework of the bounded-difference stability

根据有界差分条件,提出了学习算法的有界差分稳定框架.依据新框架,研究了机器学习阈值选择算法,再生核Hilbert空间中的正则化学习算法,Ranking学习算法和Bagging算法,证明了对应学习算法的有界差分稳定性.所获结果断言了这些算法均具有有界差分稳定性,从而为这些算法的应用奠定了理论基础.

The bounded-difference stability framework of the learning algorithms is proposed in this paper based on the bounded difference condition.Under the new framework,the machine learning threshold selection algorithm,regularization algorithms in the reproducing kernel Hilbert space （RKHS）,the Ranking algorithm and the Bagging algorithm are studied,and the bounded-difference stability properties of these learning algorithms are proved in this paper.The results show that these algorithms have the bounded-difference stability properties,and thus establish the theoretical principles for the applications of these algorithms.