邻域嵌入的张量学习

发布时间：2018-07-08 08:07

本文选题：判别邻域嵌入(DNE) + 张量子空间分析(TSA)　；参考：《计算机科学与探索》2017年07期

【摘要】：传统的机器学习算法把数据表示成向量的形式进行处理,而现实世界许多应用中的数据都是以张量形式存在的,如图像、视频数据等,如果将这些本质上非向量形式的数据强制转换成向量表示,不仅会产生维数灾难和和小样本问题,而且会破坏数据本身的内部空间排列结构,不利于发现数据的好的低维表示。判别邻域嵌入(discriminant neighborhood embedding,DNE)是比较流行的面向向量的判别分析方法,在改进DNE算法的基础上,提出了面向张量数据的局部一致保持的邻域嵌入张量判别学习(neighborhood-embedded tensor learning,NTL)算法。NTL算法不仅克服了DNE面向向量的缺点,而且弥补了DNE方法偏重数据的邻域点而忽略数据的非邻域点影响的不足,通过精心设计目标函数(嵌入3个图:同类结点的邻接图、不同类结点的邻接图、其他结点的关联图),使投影空间的同类结点更加紧凑,不同类结点更加疏远,从而增强了算法的判别能力。3个公开数据库(ORL、PIE和COIL20)上的实验验证了NTL拥有更高的识别率,同时也拥有更高的算法效率。
[Abstract]:Traditional machine learning algorithms represent data as vectors, but in the real world, many applications of data exist in the form of Zhang Liang, such as images, video data, etc. If these essentially non-vector forms of data are forced to be transformed into vector representations, not only will the dimension disaster and small sample problems occur, but also the internal spatial arrangement of the data itself will be destroyed, which is not conducive to the discovery of good low-dimensional representations of data. Discriminant neighborhood embedding (discriminant neighborhood embedding DNE) is a popular vector-oriented discriminant analysis method, which is based on the improved DNE algorithm. In this paper, a locally consistent preserving neighborhood embedded Zhang Liang discriminant learning (neighborhood-embedded tensor) algorithm for Zhang Liang data is proposed. The algorithm not only overcomes the shortcoming of vector oriented, but also overcomes the shortcoming of DNE-oriented algorithm. Moreover, the DNE method emphasizes the neighborhood point of the data and neglects the influence of the non-neighbor point of the data. By carefully designing the objective function (embedding three graphs: the adjacent graph of the same kind of node, the adjacent graph of the different class node, the adjacent graph of different classes of nodes), The association graph of other nodes makes the similar nodes in projection space more compact and the different class nodes more estranged, thus enhancing the discriminant ability of the algorithm. Experiments on three open databases (ORLPIE and COIL20) show that NTL has a higher recognition rate. At the same time, it also has higher algorithm efficiency.
【作者单位】：苏州大学计算机学院;江苏师范大学计算机学院;
【基金】：国家自然科学基金Nos.61033013,61402207,61272297~~
【分类号】：TP181

【参考文献】