基于熵权距离的图正则局部线性嵌入算法

Graph Regular Local Linear Embedding Algorithm Based on Entropy Weight Distance

针对局部线性嵌入(Locally Linear Embedding,LLE)算法在挖掘数据结构时未考虑特征权重且仅局限于数据的线性拟合关系,导致特征提取效果不佳的问题,提出一种基于熵权距离的图正则局部线性嵌入(Graph Regular Local Linear Embedding Algorithm Based on Entropy Weight Distance,EWD-GLLE)算法。首先,采用信息熵加权的余弦距离划分样本邻域,减小不重要特征对邻域划分的影响,提高了邻域划分的准确性;然后,利用融合热核权重与余弦权重的拉普拉斯图约束低维嵌入,以保留更多的原始数据信息,进而提取到更显著的特征。在两种轴承数据集上的实验结果表明:EWD-GLLE算法的特征提取性能明显优于LLE、LTSA、LDA算法。

Considering the fact that the locally linear embedding(LLE) algorithm does’ t consider the feature weight and is limited to the linear fitting relationship of data when mining the data structure and it results in poor effect in the feature extraction, a graph regular local linear embedding algorithm based on e-ntropy weight distance(EWD-GLLE) algorithm was proposed. Firstly, it has the cosine distance which weighted by information entropy adopted to divide the sample neighborhood so as to reduce the influence of unimportant features on the neighborhood division and improve its accuracy thereof;then, it has the Laplacian graph which combining the thermal kernel weight and cosine weight employed to constrain the low-dimensional embedding and to retain more original data information and extract more significant fea-tures. Experiments on two bearing data sets show that, the feature extraction performance of EWD-GLLE algorithm outperforms LLE,LTSA and LDA algorithms significantly.