一种基于非负低秩稀疏图的半监督学习改进算法

发布时间：2018-03-30 13:35

本文选题：半监督学习　切入点：图模型　出处：《电子与信息学报》2017年04期

【摘要】：该文针对基于非负低秩稀疏图的半监督学习算法不能准确地描述数据结构的问题,提出一种融合平滑低秩表示和加权稀疏约束的改进算法。该算法分别对经典算法的低秩项和稀疏项进行改进,准确地捕获了数据的全局子空间结构和局部线性结构。在构建目标函数时,使用对数行列式函数代替核范数平滑地估计秩函数,同时利用形状交互信息和有标签样本的类别信息构造加权稀疏约束正则项。然后通过带有自适应惩罚的线性交替方向方法求解目标函数并采用有效的后处理方法重构数据的图结构,最后利用基于局部和全局一致性的半监督分类框架完成学习任务。在ORL库,Extended Yale B库和USPS库上的实验结果表明,该改进算法提高了半监督学习的准确率。
[Abstract]:In this paper, the semi-supervised learning algorithm based on non-negative low-rank sparse graph can not accurately describe the data structure. An improved algorithm combining smooth low rank representation and weighted sparse constraint is proposed, which improves the low rank and sparse terms of classical algorithms, respectively. The global subspace structure and local linear structure of the data are accurately captured. In constructing the objective function, the rank function is estimated smoothly by using the logarithmic determinant function instead of the kernel norm. The weighted sparse constrained regulars are constructed by using both the shape interaction information and the class information of labeled samples. Then the objective function is solved by linear alternating direction method with adaptive penalty and an effective post-processing method is adopted. Reconstructing the graph structure of the data, Finally, the semi-supervised classification framework based on local and global consistency is used to complete the learning tasks. The experimental results on ORL libraries extended Yale B and USPS show that the improved algorithm improves the accuracy of semi-supervised learning.
【作者单位】：南京理工大学计算机科学与工程学院;
【基金】：国家自然科学基金(61473154)~~
【分类号】：TP181

【相似文献】