基于L2,1范数和流形正则项的半监督谱聚类算法

Semi-supervised spectral clustering algorithm based on L norm and manifold regularization terms

谱聚类算法受到相似矩阵的影响以及没有使用先验信息,使得聚类结果有很大的局限性。针对这一问题,提出了一种基于L2,1范数和流形正则项的半监督谱聚类算法。一方面借助L2,1范数的鲁棒性学习到合理的相似矩阵;另一方面充分利用监督信息,不仅指导了初始相似矩阵的构造,而且引入流形正则项去调整模型,从而改善聚类效果。实验结果表明,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。

The spectral clustering algorithm is affected by the similarity matrix and not using prior information, which makes the clustering results with great limitations. For this problem, we propose a semi-supervised spectral clustering algorithm based on L2,1 norm and manifold regularization terms. With the help of robustness in L2,1 norms, a reasonable similarity matrix is learned. In addition, full use of supervisory information not only is added in the initial similarity matrix, but also is used in manifold regularization term to adjust the model, thereby improving the clustering effect. The clustering results of the proposed clustering algorithm on artificial data sets and real data sets are more effective than other clustering algorithms in most cases.