基于图稀疏的自表达属性选择算法

Graph self-representation for sparse feature selection

为解决高维数据属性维度高,不易直接应用的问题,提出通过属性自表达移除不相关和冗余属性的属性选择算法。基于稀疏学习的框架,通过属性自表达考虑属性间的相关性,利用子空间学习的局部保留投影(LPP)算法,确保属性选择时数据的局部结构保持不变。实验结果表明,该算法在UCI等数据集上优于4种对比算法。

To solve the issues that high-dimensional data are hardly used in applications,a feature selection method using the self-representation of samples to remove the redundant and irrelevant features was proposed.The self-representation of samples was used to estimate the correlation among features in a sparse feature selection framework,and also locality preserving projection（LPP）was employed to preserve the local structures of samples.The experimental results on real datasets show that the proposed method outperforms state-of-the-art methods.