【作者】 张召； 业宁； 业巧林；

【Author】 ZHANG Zhao~1,YE Ning~(1,2),YE Qiao-lin~3 1 School of Information Technology,Nanjing Forestry University,Nanjing,210037,China 2 School of Computer Science and Technology,Shandong University,Jinan,250100,China 3 Department of Computer Science,Nanjing University of Science & Technology,Nanjing 210094,China

【机构】 南京林业大学信息科学技术学院； 山东大学,计算机科学与技术学院； 南京理工大学计算机系；

【摘要】 降维是在损失较少信息的情况下处理高维图像数据的关键技术,是高维数据预处理的重要步骤。本文研究了基于配对约束和混合核函数的半监督非线性降维方法KS~2DR该方法可有效利用标签和未标签的样本执行半监督学习。基于配对约束形式的领域知识被用于判断当前样例是属于相同类(相似约束)还是不同类(不相似约束)。KS~2DR先将样本数据投影成一系列"有用的"特征形式,同时保持原始数据以及定义在投影后的低维特征空间中的相似约束与不相似约束下样本的内部结构特征,使样本在投影空间中容易有效地区分。通过大量的分类任务和数据可视化研究,结果表明,当样本数据的维数被减少到一个较低的水平时,KS~2DR的性能表现的更为优越,几乎总是取得较高的分类精确,随着投影维数的不断变化,算法体现出广泛的适应性和良好的学习能力。在相同的测试条件下,KS~2DR算法运行花费较少时间,其整体性能甚至优于典型的PCA,KPCA和KFD方法。更多还原

【Abstract】 Dimensionality reduction is one of key interests in dealing with high-dimensional image data without losing intrinsic information,which is an important preprocessing step in high-dimensional data analysis.The problem of semi-supervised dimensionality reduction with kernels called KS~2DR is considered to exploit the unlabeled samples.In this setting,domain knowledge in the form of pairwise constraints is adopted to specify whether pairs of instances belong to the same class or different classes and a mixture kernel function is designed to improve the performance ofKS~2DR.KS~2DR can preserve the intrinsic structure of the unlabeled data as well as both the similar and dissimilar pairs constraints defined on the labeled training samples in the projected low-dimensional kernel space,under which the samples are easier to be effectively partitioned from each other. We demonstrate the practical usefulness and high scalability of KS~2DR method in data visualization and classification tasks through extensive simulation studies.Experimental results show KS~2DR can almost always achieve the highest accuracy when the dimensionality is reduced to a very low level.The performance of KS~2DR outperform those of some established typical dimensionality reduction methods no matter how many number of constraints,dimensions and kernel tunable parameter are used.更多还原