一种半监督判别邻域嵌入算法

Semi supervised discriminant neighborhood embedding algorithm

邻域保持嵌入(Neighborhood Preserving Embedding,NPE),作为局部线性嵌入(Locally Linear Embedding,LLE)的线性化版本,由于在映射前后保持了数据的局部几何结构并得到了原始数据的子空间描述,在模式识别领域具有较强的应用价值。但作为非监督处理算法,在具体的模式分类中有一定局限性,提出一种NPE的改进算法——半监督判别邻域嵌入(SSDNE)算法,引入标记后样本点的类别信息,并在正则项中引入样本的流形结构,最大化标记样本点的类间信息和类内信息。既增加了算法的辨别能力又减少了监督算法中对样本点进行全标记的工作量。在ORL和YaleB人脸库上的实验结果表明,改进的算法较PCA、LDA、LPP以及原保持近邻判别嵌入算法的识别性能有了较明显的改善。

As a linear version of Locally Linear Embedding（LLE）,Neighborhood Preserving Embedding（NPE） can preserve the local geometric structure after embedding and obtain the subspace of the original dataset,hence it has wide practical application in pattern recognition.But it is an unsupervised learning algorithm,which gives rise to some limitations in pattern classification.In this paper,a Semi Supervised Discriminant Neighborhood Embedding（SSDNE） based on NPE is proposed,in which the class information of the labeled samples is considered,also incorporates a regularizer which takes into account of the whole manifold structure.Finally,the algorithm maximizes the ratio of interclass scatter to the intra-class scatter.So the discriminated ability is enhanced and complexity of all labeled data as the case in supervised learning is reduced.Numerical experiments on the ORL and YaleB face databases show that the proposed method outperforms the PCA,LDA,LPP and original NPE algorithm.