基于稀疏流形聚类嵌入模型和L1范数正则化的标签错误检测

综合利用含错标签中的有用信息和数据结构中蕴含的鉴别信息,提出一种基于稀疏流形聚类嵌入模型和L1范数正则化的标签错误检测修正方法.首先,用稀疏流形聚类嵌入模型将数据投影到易分类的空间,利用标注正确的极少量样本和最近邻分类器获得新标签;然后,构造标签错误检测模型,获得仅含0、1元素的检测向量,正确、错误的标签分别对应着1、0的位置;最后,给出了相应的优化算法及收敛证明,并在相关实验上验证了算法的有效性.

Labeling errors detecting and correcting algorithm based on sparse manifold clustering and embedding and L1 norm regularization

As to detect and correct the labeling errors, a labeling errors detecting and correcting algorithm based on sparse manifold clustering and embedding and L₁norm regularization is proposed. The proposed algorithm is based on the useful information in the original labels and the natural discriminating information which is contained in the data structure. Firstly, the original data are projected to the new space by using the sparse manifold clustering and embedding model. Then, a nearest neighbor classifier and a very small amount samples which are labeled correctly are used to obtain new labels for the original data. Meanwhile, the constructing labeling error detection model is built and then the sparse label detection vector which consists of 0 and 1 is obtained to modify the detection errors. The inaccurate and accurate labels correspond to 0 and 1 in the label detection vector respectively. Finally, the convex optimization scheme is introduced to solve the optimization problem and the convergence proofs are given. The experiment results show the effectiveness of the proposed algorithm based on the artificial data of complex manifold structure and the typical low-dimensional, high-dimensional data.