计算机技术与发展2019，Vol.29Issue(3)：6-11,6.DOI:10.3969/j.issn.1673-629X.2019.03.002
基于结构化稀疏投影的多视图特征提取框架
A Framework of Multi-view Feature Extraction Based on Multiple Structured Sparsity Projection
摘要
Abstract
Multi-view learning is usually based on two important principles:consistency and complementarity. Consistency derives from the shared information between views and complementarity derives from the unique information of different views. However, the existing multi-view learning algorithms usually focus on only one of them, so that the resulting model performance is not optimal. For this, we propose a new multi-view feature extraction algorithm framework called multiple structured sparsity projection (MSSP), which can simultaneously extract the common and unique information of the view. The objective function of MSSP contains two parts:fusion projection discriminant and structured sparsity regularization. Then the consistent and complementary information of different views can be fused simultaneously. In addition, we also give an algorithm based on the gradient decent over the Stiefel manifold to obtain a local optimal solution, and analyze how the hyper-parameter change with the dimension of the projection. The experiment verifies the effectiveness of MSSP and the trend of the hyper-parameter.关键词
一致性/互补性/多视图学习/特征提取/结构稀疏/Stiefel流形Key words
consistency/complementarity/multi-view learning/feature extraction/structured sparsity/Stiefel manifold分类
信息技术与安全科学引用本文复制引用
江帆,田青..基于结构化稀疏投影的多视图特征提取框架[J].计算机技术与发展,2019,29(3):6-11,6.基金项目
国家自然科学基金(61702273) （61702273）
江苏省自然科学基金(BK20170956) （BK20170956）
