摘要
二类分类问题是机器学习中的最基本的一类重要问题。目前广泛使用的,也是最为有效的学习算法是支持向量机(SVM)。然而对于某些非线性分类问题,SVM还不能给出令人满意的解,因此希望能找到一种方法对SVM解决非线性分类问题的能力加以改进。对二类分类问题,提出一种基于感知器的样本空间划分方法。该方法首先用感知器提取样本的分布信息,将整体问题划分为局部空间中的分类问题,而后使用SVM求出各个局部问题的最优分界面,并用最小最大模块化网络对局部分界面进行综合,得到问题的全局解。仿真实验表明,新方法能够有效地分析样本空间,提取样本分布信息,在测试数据上得到了比原有方法更好的准确率。新方法实现了预期的目标,提高了分类器处理非线性分类问题的能力。
Binary classification is a fundamental problem in machine learning area. Currently, the widely used and best performing learning method is support vector machine (SVM). However, SVM cannot give satisfactory answer to some non-linear problems. A new approach to improve SVM's ability on non-linear problems is expected. Focusing on binary classification problems, a novel sample space analyzing method based on perceptron is proposed. The method starts from extracting sample distribution information, and divides the overall problem into a series of local problems. Then finding optimized local separator with SVM, and finally combining the local separators with the minimization and the maximization principles to get the overall classifier. The simulation results indicate that the proposed method can effectively analyze sample space and extract distribution information, and achieve better prediction accuracy than existing methods. The new method can meet the requirement, and improve classifier's performance on non-linear problems.
出处
《计算机仿真》
CSCD
2008年第2期96-99,103,共5页
Computer Simulation
基金
国家自然科学基金(60375022
60473040)
上海交通大学微软智能计算与智能系统实验室的资助
关键词
感知器
支持向量机
模式识别
样本空间分析
最小最大模块化网络
Perceptron
Support vector machine (SVM)
Pattern recognition
Sample space partition
Min-max modular network