一种半监督稀疏保持近邻判别嵌入算法

Semi-supervised Sparse Neighborhood Preserving Discriminant Embedding Algorithm

保持近邻嵌入(NPE)算法对局部线性嵌入(LLE)算法进行了改进,克服了新来样本问题,但在处理分类问题上表现不足。基于此提出一种半监督稀疏保持近邻判别嵌入算法,该方法首先采用小波变换对数据进行预处理,然后执行等距离映射(Isomap)算法选择合适的低维嵌入维数,最后结合稀疏表示理论、NPE和线性判别分析(LDA)的思想,重构邻域图,并在建立目标函数时使得已标签信息中同类样本点之间相互靠近,异类样本点之间相互远离,未标签信息邻域信息得以保持。这样,既得到了高维映射函数,又提高了分类正确率。通过在人脸数据库上实验,并与其他半监督算法作比较,该算法在识别率上表现较好。

Neighborhood preserving embedding（ NPE ）algorithm is improved on Locally Linear Embedding （LLE）algorithm, and it overcomes the new coming sample problem,but it is not good in dealing with the classification. A semi-supervised sparse neighborhood preserving diseriminant embedding algorithm is presented,the method preprocesses the data by using the wavelet transform, and then it performs Isomap algorithm to select the appropriate low-dimensional embedding dimension,and the last it reconstructs the neighborhood graph which is based on the theroy of sparse representations, NPE and linear diseriminant analysis（LDA） , at the same times, it makes closer between the same class points and away from each other between different class points which have been labeled, maintains the information of points which have been unlabeled. So the new algorithm has both got the high-dimensional mapping function, and it improves the classification accuracy. Experiments on face databases, comparing with other semi-supervised algorithms, the pro- posed algorithm performes better on the recognition rate.