基于自适应相似性的无监督多视图特征选择

Unsupervised multi-view feature selection based on adaptive similarity

为解决传统特征选择方法忽略视图内部特征的相关性及不同视图之间的特征关联性问题,提出一种基于自适应相似性的特征选择学习方法。在特征选择时考虑视图内部的特征相关性,对每个视图进行特征选择,通过引入图正则化,充分利用数据的局部几何特性,使同类别特征之间的联系更加紧密,达到增强算法的鲁棒性。引入L_(1/2)稀疏范数降低噪声,提高分类模型的准确率。通过与现有的特征方法进行对比分析,提出方法在ACC和NMI上优于其它方法。

To solve the problem that the traditional feature selection method ignores the correlation between the internal features of the view and the feature correlation between different views,a feature selection learning method based on adaptive similarity was proposed.The internal characteristics of the view were considered when selecting features.Feature relevance and feature selection were implemented for each view.At the same time,graph regularization was introduced to make full use of the local geometric characteristics of the data,the relationship between features of the same category was then closer,and the robustness of the algorithm was enhanced.The introduction of the L_(1/2) sparse norm effectively reduced the noise while improving the accuracy of the classification model.Through comparative analysis with the existing feature methods,the proposed method is superior to other methods in ACC and NMI.