最优间隔分布脊回归

发布时间：2018-01-27 02:20

本文关键词： 脊回归流形正则化最优间隔分布间隔方差全局结构　出处：《计算机研究与发展》2017年08期 　论文类型：期刊论文

【摘要】：脊回归(ridge regression,RR)是经典的机器学习算法之一,广泛应用于人脸识别、基因工程等诸多领域.其具有优化目标凸、存在闭合解、可解释性强以及易于核化等优点,但是脊回归的优化目标并没有考虑样本之间的结构关系.监督流形正则化学习是最具代表性的、最成功的脊回归正则化方法之一,其通过最小化每类类内方差来考虑样本之间的类内结构关系,可是单纯地只考虑类内结构仍然不够全面.以一种全新的视角重新审视最近提出的"最优间隔分布学习"原理,发现了最优间隔分布的目标可以同时优化类内间隔方差和类间间隔方差,从而同时优化了局部的类内结构和全局的类间结构.基于此提出了一种充分考虑数据结构化特征的脊回归算法——最优间隔分布脊回归(optimal margin distributionmachine ridge regression,ODMRR)算法,该算法具有RR以及MRRR(manifold regularization ridge regression)的各种优势.最后通过实验验证了该方法具有优越的性能.
[Abstract]:Ridge regression (RRR) is one of the classical machine learning algorithms, which is widely used in face recognition, genetic engineering and many other fields. There are some advantages such as closed solution, strong interpretability and easy to nucleate, but the optimization target of ridge regression does not take into account the structural relationship between samples. The supervised manifold regularization learning is the most representative. One of the most successful methods of ridge regression regularization is to minimize the intra-class variance of each class to consider the intra-class structural relationships between samples. However, it is still not comprehensive to consider only the intra-class structure. The recently proposed principle of "optimal interval distribution learning" is re-examined from a new perspective. It is found that the objective of optimal interval distribution can optimize both intra-class and inter-class interval variances. Therefore, the local intra-class structure and the global inter-class structure are optimized at the same time. Based on this, a ridge regression algorithm with optimal interval distribution is proposed, which takes full account of the structured features of the data. Optimal margin distributionmachine ridge regression. ODMRR algorithm. The algorithm has RR and MRRR(manifold regularization ridge regression. Finally, the advantages of this method are verified by experiments.
【作者单位】：计算机软件新技术国家重点实验室(南京大学);软件新技术与产业化协同创新中心;
【基金】：国家自然科学基金项目(61673201)~~
【分类号】：TP181
【正文快照】： (1)此类方法以往被称为“大间隔分布学习”方法[17-18],但由于“间隔分布”本身并没有“大小”之别,因此“最优间隔分布”这个称谓比“最大间隔分布”更合适[16].脊回归(ridge regression,RR)[1]的优化目标致力于最小化模型输出以及样本标记之间的二次误差,在人脸识别[2]、基

【参考文献】