光学 精密工程, 2014, 22 (7): 1921, 网络出版: 2014-09-01
基于监督保局子空间虚假近邻准则的原始特征选择
Original feature selection based on false nearest neighbor criterion in supervised locality preserving subspace
监督保局投影 虚假近邻点 特征选择 模式分类 低阻油气层识别 Supervised Locality Preserving Projection(SLPP) False Nearest Neighbor(FNN) feature selection pattern classification low resistivity hydrocarbon reservoir recognition
摘要
提出一种基于监督保局投影(SLPP)与虚假最近邻(FNN)准则的原始特征选择方法。该方法首先将非线性原始数据映射到监督保局子空间, 消除样本数据输入变量之间的相关性; 然后, 利用虚假近邻点方法计算剔除每个原始特征前后输入样本在监督保局子空间里的相似性测度, 获得每个原始特征对类别变量不同程度的解释力; 最后, 从全特征开始逐步剔除解释能力弱的特征进而获得多组特征子集, 并建立最近邻分类器, 识别率最高且含特征数最少的特征子集即为最优特征子集。采用合成数据对该方法进行了仿真验证, 结果表明, 该方法可获得与数据集本质分类特征吻合的最佳特征子集。将该方法应用于选择真实的低阻油气层特征, 获得的最佳特征子集比全特征集合的特征数量减少了50%以上, 分类识别率高出8%。结果显示该方法具有优秀的原始特征选择能力, 是一种有效的非线性特征选择方法。
Abstract
A novel method based on Supervised Locality Preserving Projection (SLPP) and False Nearest Neighbor (FNN) was proposed for selecting the most proper feature for nonlinear pattern classification.In the proposed method, nonlinear original data were mapped to the supervised locality preserving subspace to eliminate the existing multi-collinearity among the features. Then, the interpretation capability for original features was estimated through calculating the variable mapping distance in the supervised locality preserving subspace. The nearest neighbor classifier based on each subset obtained by eliminating weak features successively was constructed. Finally, the optimal feature subset was selected corresponding to the highest recognition accuracy and the least number of features.The experiment on synthetic dataset shows that the proposed method can obtain an optimal feature subset containing the essential features in accordance with the classification goal. The method was used to select the features of low resistivity hydrocarbon reservoir, and the result indicates that the obtained optimal feature subset contains over 50% less feature and achieves 8% higher recognition accuracy as compared to that of the all-feature set. These results validate that the proposed method can offer excellent abilities of original feature selection and nonlinear feature selection.
辜小花, 李太福, 杨利平, 易军, 周伟. 基于监督保局子空间虚假近邻准则的原始特征选择[J]. 光学 精密工程, 2014, 22(7): 1921. GU Xiao-hua, LI Tai-fu, YANG Li-ping, YI Jun, ZHOU Wei. Original feature selection based on false nearest neighbor criterion in supervised locality preserving subspace[J]. Optics and Precision Engineering, 2014, 22(7): 1921.