摘要
针对随机近邻嵌入算法的非线性本质和无监督学习特征,提出一种线性有监督的特征提取方法,称为判别随机近邻嵌入分析.该方法通过输入样本的类别信息构建数据分布的联合概率表达式,用于反映同类和异类数据间的相似度;同时引入线性投影矩阵生成子空间数据,并在类内KL散度最小和类间KL散度最大的准则下建立目标泛函.通过人工合成数据和经典人脸库对文中方法的性能进行验证,结果表明,该方法不仅具有较好的可视化能力,而且能够有效地对不同类别的数据进行降维分簇,提升后续模式分类器的鉴别效果.
A novel linear supervised feature extraction method named discriminative stochastic neighbor embedding （DSNE） is proposed based on the algorithm of stochastic neighbor embedding that is unsupervised and nonlinear. DSNE selects the joint probability to model the pairwise similarities of input samples with class labels. The linear projection matrix is used to discover the underlying structure of data manifold which is nonlinear. The cost function is constructed to minimize the intraclass Kullback-Leibler divergence as well as maximize the interclass Kullback-Leibler divergences. DSNE is evaluated in artificial synthetic data and face database. Experimental results suggest that the proposed algorithm provides a better visualization effectiveness as well as powerful pattern revealing capability for complex manifold data.
出处
《计算机辅助设计与图形学学报》
EI
CSCD
北大核心
2012年第11期1477-1484,共8页
Journal of Computer-Aided Design & Computer Graphics
基金
国家自然科学基金(61070043)
浙江省自然科学基金(LQ12F03011)
关键词
流形嵌入
有监督学习
数据可视化
随机近邻嵌入
manifold embedding
supervised learning
data visualization
stochastic neighbor embedding
