基于L21范数和回归正则项的半监督聚类算法

Semi-supervised Clustering Algorithm Based on L21 Norm and Regression Regular Term

针对半监督学习中基于线性嵌入的回归正则项难以捕获数据流形结构的问题,提出基于L21范数和回归模型的半监督聚类算法。一方面充分利用监督信息,指导初始相似矩阵的构造,并利用L21正则项构造标签矩阵F的弹性嵌入回归模型;另一方面借助L21范数的鲁棒性学习合理的相似矩阵,从而改善聚类效果。通过实验表明,所提出的聚类算法在人工数据集和真实数据集上的聚类结果较其他聚类算法更加有效。

Linear embedding based regression model was difficult to capture data manifold structure in semi-supervised learning.To solve the problem a semi-supervised clustering algorithm based on L21 norm and regression regular term was proposed.On one hand,the supervision information was fully used to guide the construction of initial similarity matrix,and the elastic embedded regression model of label matrix F was constructed by using L21 regular terms.On the other hand,by virtue of the robustness of L21 norm,a reasonable similarity matrix was learned to improve the clustering effect.Experiments showed that the proposed clustering algorithm was more effective than other clustering algorithms on artificial datasets and real datasets.