基于模糊核匹配追寻的特征模式识别

具体的实验结果。由于采用的样本为非平衡样本（即两类样本个数相差较大），传统的核匹配追寻不能对弱势样本（数量小的一类样本）进行有效的识别，甚至失去了识别能力（识别率<50％），而采用模糊核匹配追寻，就可以有效地解决这一问题，仿真的试验结果中，当标准KMP对弱势样本的识别率<50％（34.78％）时，利用模糊KMP，仍然可以使识别精度可以达到99％以上。

表1：对UCI非平衡数据的测试

数 据 训练样本 检验样本 损失函数 算法 Fuzzy KMP Loss-mse Breast Cancer ?1类：58 ?1类：142 ?1类：23 5支撑模式 4 55 6 64 87 100 69 102 50 50 48 50 识别率 99.92％ 34.78％ 99.87% 33.56% 99.14％ 56.32％ 99.25% 57.22% 100％ 84.62％ 100% 83.69% KMP Fuzzy KMP Loss-tanh KMP Fuzzy KMP Loss-mse 6 Pima Indians Diabetes ?1类：94 ?1类：174 KMP Fuzzy KMP Loss-tanh KMP Fuzzy KMP Loss-mse ?1类：162 ?1类：39 KMP ?1类：26 Thyroid ?1类：101 Fuzzy KMP Loss-tanh KMP 5、 总结

核匹配追寻具有很强的推广能力，强大的非线性处理能力和高维处理能力，同时较其他核机器相比，其稀疏性更优。然而在实际问题中经常遇到这样几种情况：1）所获得的样本是具有时间属性的；2）要求其中一类样本的识别精度；3）非平衡样本的识别。由于传统的核匹配追寻在处理模式识别的问题平等对待所有的样本，它要求总识别误差尽可能的小，但是并不能对某一类或某一些指定的样本进行针对性的识别，这就限制了核匹配追寻在这些实际问题中的应用。

针对这些问题，本文提出了模糊核匹配追寻，根据问题的要求对每个样本作出重要性定义，学习机

56

Loss-mse表示核匹配追寻采用平方损失函数。

Loss-tanh表示核匹配追寻采用修正双曲正切损失函数。

13

可以根据样本的重要性定义进行程度不同的学习，对次要的样本粗略学习，而对重要的样本进行充分学习，使学习机的最终判决对指定的重要样本达到较高的识别精度。本文进行了大量的仿真实验，结合分类图例证实了模糊核匹配追寻可行性及有效性；在对UCI数据的性能测试中可以得出：当传统的核匹配追寻已不能对弱势样本进行识别（识别率小于50％）时，模糊核匹配追寻仍然对弱势样本保持了较高的识别精度。

参考文献

[1] [2]

Pascal Vincent, Yoshua Bengio. Kernel matching pursuit. Machine Learning, 48:165--187, 2002. Mallat S., Z. Zhang (1993, Dec.). Matching pursuit with time-frequency dictionaries. IEEE Trans. Signal Proc. 41 (12), 3397-3415.

[3]

Davis G., Mallat S., Z. Zhang. Adaptive time-frequency decompositions. Optical Engineering 33(7), 2183-2191.

[4]

S. Mallat. A theory for nuliresolution signal decomposition: The wavelet representation. IEEE Trans. Pattern Anal. Machine Intell., vol. 11, pp. 674-693, July 1989.

[5]

Pati Y., Rezaiifar R. and Krishnaprasad P. Orthogonal Matching Pursuit: Recursive Function Approximation with Applications to Wavelet Decomposition. In Proceedings of the 27th Annual Asilomar Conference on Signals, Systems, and Computers, pp. 40-44, 1993.

[6]

Vapnik, V.N. An overview of statistical learning theory, IEEE Trans. Neural Networks, vol. 10, no.5, pp.988-999, 1999.

[7]

C. J. C. Burges. A tutorial on support vector machines for pattern recognition.. Data Mining and Knowledge Discovery, 1998, 2(2): 1-47.

[8] [9]

Sch?lkopf, B., Smola, A. Learning with kernels. MIT Press, 1999.

Burges, C. J. Geometry and invariance in kernel based method. In advance in kernel method-Support vector learning. Cambridge, MA: MIT Press, 1999, pp. 86-116.

[10]

Cao, L. J. and Francis, E. H. Support vector machine with adaptive parameters in financial time series forecasting. IEEE Trans. Neural Networks, vol. 14, no. 6, pp. 1506-1518, 2003.

14

Pattern Recognition Based on the Fuzzy Kernel Matching Pursuit

LI Qing, JIAO Li-cheng, Zhou Weida

( Institute of Intelligent Information Processing, Xidian University, Xi’an 710071 )

Abstract Kernel Matching Pursuit (KMP), a novel method of the pattern recognition, presents excellent performance in solving the problems with small sample, nonlinear and local minima. KMP has been proposed to provide a good generalization performance for both classes, yet the classification precision of some important data can’t be classified precisely. This is mainly because the decision function found by KMP is the synthetic consideration results of all the data, which has greatly limited its use in many practical problems, such as time series identification and unbalanced data classification. In this paper, an fuzzy kernel matching pursuit machine is (FKMP) proposed, which can classify the appointed important samples much more precisely according to the predefined importance of the data. Lots of experiments have been done in the paper to prove the feasibility and validation of the fuzzy kernel matching pursuit machine.

Keywords Machine Learning, Kernel Matching Pursuit, Fuzzy Kernel Matching Pursuit; Time Series Identification; Unbalanced Data Classification

15

? Background

We have made researches on many fields of the support vector machine, such as Linear programming support vector machine, kernel matching pursuit classifier ensemble, support vector regression based on unconstrained convex quadratic programming and so on, and have applied these methods to the field of SAR image processing, recognition of plane HRRP and many other fields. This paper belongs to the part of novel method of machine learning and focuses on proposing an fuzzy kernel matching pursuit machine (FKMP), which can classify the appointed important samples much more precisely according to the predefined importance of the data, so as to develop the practical applications of the KMP.

? 作 者 简 历

李 青，男，1979年生，博士，工程师，研究方向为机器学习、模式识别及统计学习理论。

LI Qing, male, born in 1979, Ph.D.. His current research interests include machine learning, pattern recognition and statistic learning theory.

焦李成，男，1959年生，博士，教授，博士生导师，研究方向为非线性理论、神经网络、数据挖掘、进化算法与子波理论。

JIAO Licheng, male, born in 1959, Ph.D., professor and Ph.D. supervisor. His current research interests include nonlinear theory, neural network, data mining, evolutionary computation and wavelet theory.

周伟达，男，1974年生，博士，副教授，主要研究领域包括机器学习、模式识别等。

ZHOU Weida, male, born in 1974, Ph.D.. His current research interests include intelligent information processing, machine learning, statistic learning theory and data mining.

联 系 人：李 青

电 话：13951943350，025-83773124（办公室） 通信地址：南京1313信箱100分箱 邮政编码：210013

E-mail ：kingdomyangfan@hotmail.com

16