多核学习中基于复合梯度映射的学习算法研究.pdfVIP

优先出版 计 算 机 应 用 研 究 第32 卷 多核学习中基于复合梯度映射的学习算法研究 a b 龙文光 ，刘益和 ( 内江师范学院 a．现代教育技术中心；b ．计算机科学学院，四川 内江，641112) 摘 要：多核学习技术在机器学习、数据挖掘等领域具有重要应用。现有的多核学习算法大多假设训练样本分类完全正 确，将其应用到受扰分类样本上时，由于分类存在差错，因此往往只能实现次优性能。为了解决这一问题，文中首先将 受扰分类多核学习问题建模为随机规划问题，并得到一种极小极大表达式，然后提出基于复合梯度映射的一阶学习算法 对问题进行求解，理论分析表明，本文算法的收敛速度为O(1/T)，大大快于传统算法的收敛速度 O(1/√T) 。最后，基于 5 个UCI 数据集的实验结果也验证了本文观点和优化算法的有效性。 关键词：多核学习；训练样本；随机规划；复合梯度映射；收敛速度；UCI 数据集 中图分类号：TP391 文献标志码：A Research on learning algorithm based on composite gradient mapping in multiple kernel learning 1 2 LONG Wen-guang , LIU Yi-he (a. Modern Educational Technology Center, b. College of Computer Science, Neijiang Normal University, Neijiang, Sichuan 641112, China) Abstract: Multiple kernel learning technology has important application in the fields of both machine learning and data mining, the existing multiple kernel learning algorithms assume that the training sample classification entirely correct, when they are used to the classification of disturbed samples, the suboptimal performance can only be achieved due to the incorrect class assignments. In order to solve this problem, firstly, the multiple kernel learning problems from noisy labels are modeled into a stochastic programming problem, and presenting a min-max formulation, and then a first order learning algorithm based on composite gradient mapping is proposed to solve this problem. The theoretical analysis shows tha