基于高维空间最小生成树自适应覆盖模型的可拒绝分类算法
doi: 10.3724/SP.J.1146.2009.00021
A Classification Algorithm with Reject Option Based on Adaptive Minimum Spanning Tree Covering Model in High-dimensional Space
-
摘要: 在高维空间样本较少的情况下,基于统计模型的可拒绝分类方法难以对样本分布的复杂几何形体构建合理的覆盖模型。为此,该文提出基于高维空间最小生成树自适应覆盖模型的可拒绝分类模型。该模型采用最小生成树刻画高维空间样本点分布,将图形的边作为新增虚拟样本以提供更好的同类样本分布描述。通过将同类相近样本划分到一个连通几何覆盖区域内,将不同类的相近样本归于不同几何覆盖区域内,实现对不同训练类的覆盖。为了克服因不合理虚拟样本造成分类器拒识性能的下降,引入自适应调整覆盖半径策略,实现对训练类的紧致性覆盖。对于测试样本,根据训练类覆盖边界便可对其作出拒识或者接受的处理,针对交叉覆盖的接受样本,再根据数据场策略确定其真正归属类别。实验结果表明本文方法合理有效。Abstract: For small sample size problem in high-dimensional space, conventional classifiers with reject option based on statistical model could not construct appropriate covering decision boundary on data distribution. In this case, a novel adaptive Minimum Spanning Tree (MST) covering model based classifier with reject option is proposed in this paper according to the data distribution in high-dimensional space. The algorithm describes the target class using MST with the assumption that the edges of the graph are also basic elements of the classifier which offers additional virtual training data for a better coverage. By this model, similar samples from the same class are divided into a connected geometric coverage area, and similar samples from different classes are divided into different geometric coverage areas. Furthermore, in order to reduce the degradation of the rejection performance due to the existence of unreasonable additional virtual training data, an adjustable coverage radius strategy is presented in coverage construction. Then the test pattern of non-training classes could be rejected by the coverage decision boundary, and if a pattern is accepted in the cross coverage area, the recognition result is decided by the data fields model. Experiments show that the method is valid and efficient.
-
Camci Fatih and Chinnam R B. General support vector representation machine for one-class classification of non-stationary classes[J].Pattern Recognition.2008, 41(10):3021-3034[2]Lee Ki-young, Kim Dae-won, Lee Kwang-H, and Lee Doheon. Density-induced support vector data description [J].IEEE Transactions on Neural Networks.2007, 18(1):284-289[3]Guo S M, Chen L C, and Tsai J S H. A boundary method for outlier detection based on support vector domain description[J].Pattern Recognition.2009, 42(1):77-83[4]Zhu Xiao-kai and Yang De-gui. Multi-class support vector domain description for pattern recognition based on a measure of expansibility. Acta Electronica Sinica, 2009, 37(3): 464-469.[5]Wu Jian-ming. Multilayer Potts perceptrons with Levenberg- Marquardt learning[J].IEEE Transactions on Neural Networks.2008, 19(12):2032-2043[6]Kokiopoulou Effrosyni and Frossard Pasca. Minimum distance between pattern transformation manifolds: algorithm and applications[J].IEEE Transactions on Pattern Analysis and Machine Intelligence.2009, 31(7):1225-1237[7]Wright J, Yang A Y, Ganesh A, Sastry S S, and Ma Yi. Robust face recognition via sparse representation [J].IEEE Transactions on Pattern Analysis and Machine Intelligence.2009, 31(2):210-227[8]Lin Jia-xiang, Ye Dong-yi, Chen Chong-cheng, and Gao Miao-xian. Minimum spanning tree based spatial outlier mining and its applications. 3rd International Conference on Rough Sets and Knowledge Technology, RSKT 2008, Chengdu, China, 2008(5009 LNAI): 508-515.[9]Wu Li-li and Wang Shou-jue. Study on closed-set speaker identification based on biomimetic pattern recognition[J]. Chinese Journal of Electronics, 2009, 18(2): 259-261.[10]Wang Shou-jue and Liu Xing-xing. Mathematical symbols and computing methods in high dimensional biomimetic informatics and their applications[J]. Chinese Journal of Electronics, 2008, 17(1): 1-7.
计量
- 文章访问数: 3639
- HTML全文浏览量: 110
- PDF下载量: 913
- 被引次数: 0