基于测地距离与极限学习机的监督型LLE算法研究

RESEARCH ON SUPERVISED LLE ALGORITHM BASED ON GEODESIC DISTANCE AND EXTREME LEARNING MACHINE

局部线性嵌入算法LLE(Locally Linear Embedding)是一种有效的非线性降维方法,但是该算法没有考虑样本的类别标签,并且欧式距离无法精确表示非线性数据的流形结构。针对以上LLE方法的缺陷,提出一种结合测地距离与样本类别信息的监督型LLE算法(ISO-SPLLE)。首先在LLE算法的近邻选择中使用测地距离作为相似性度量,然后利用极限学习机求出其映射函数后进行分类测试。将ISO-SPLLE算法与其他改进的LLE算法在UIC标准数据集与基因数据集上进行对比实验,结果表明,该方法对已知类别的数据能更有效地进行降维与识别。

The locally linear embedding （LLE） algorithm is one of the efficient nonlinear dimensionality redaction techniques. But it does not take the category labels of the sample into account, and its Euclidean distance can notaccurately reflect the manifold structure of nonlinear data as well. In light of the above defects in LLE, the paper proposes a supervised LLE algorithm （ISO-SPLLE） which combines the geodesic distance with category information of sample. First, in near neighbourhood selection of IJ.E algorithm the geodesic distance is used as the sim- ilarity metric, then the ELM is used to find its mapping function followed by categorical test. We conduct the comparative experiment between ISO-SPLLE algorithm and other improved LLE algorithms on UIC standard dataset and gene dataset. Result shows that this one is more effec- tive in dimensionality reduction and recognition on those category-known data.