一种基于Laplacian的半监督特征选择模型

A Semi-Supervised Feature Selection Framework Based on Laplacian

针对LASSO算法及有关扩展模型忽略样本数据间关联信息的问题,以及有标签样本难以获取的实际情况,提出了一种半监督学习的特征选择模型。引入LASSO稀疏项,去除冗余特征,选择有效特征;引入Laplacian正则项,用于保留同类有标签和无标签样本内在的几何分布信息,帮助模型选出更具有判别能力的特征集;通过相似矩阵来重构半监督特征选择模型。在UCI数据集上的分类试验结果表明,这种方法能有效提高分类性能,同时也说明样本的几何分布信息是不应被忽略的。

The traditional feature selection methods often ignore the association information between sample data,and also it is difficult to obtain the labeled samples required in full supervised learning.A semi-supervised feature selection framework based on Laplacian is proposed(SLFS).The LASSO sparse term is introduced to remove redundant features and select effective features.Laplacian regularization term is introduced to preserve the intrinsic geometric distribution information of the same kind of labeled and unlabeled samples,and help obtain more discriminant features.The semi-supervised feature selection framework is reconstructed by similarity matrix.Experimental results on UCI datasets show that the proposed method can effectively improve the classification performance,and the geometric distribution information of samples should not be neglected.