L1 Graph联合转换学习模型的多观测样本分类算法

Multiple observation sets classification algorithm based on L1 Graph transfer learning unified model

不同分布多观测样本分类问题中,训练样本和测试样本来自不同的域,针对如何利用转换学习提高不同分布多观测样本分类性能问题,提出L1-Graph联合转换学习的多观测样本分类算法。首先基于转换学习构建一种非负矩阵三因子分解框架,将其中不变信息作为源域到目标域的转换桥梁;其次,基于稀疏表示思路构造L1-Graph,自适应寻找数据近邻,保留样本及特征几何结构;最后,将两个互补目标函数联合到统一优化问题中,然后利用迭代算法解决优化问题,进而估计出测试样本类别。在USPS-Binary数字数据库、Three-Domain Object Benchmark数据库和ALOI数据库上进行对比实验,实验结果表明该方法的有效性,既提高了识别精度又保证了算法鲁棒性。

In the classification problem of multiple observation sets with different distributions,the training samples and test samples are from different domains; aiming at how to use transfer learning to improve the classification performance of multiple observation sets with different distributions,a multiple observation sets classification algorithm based on L1-Graph transfer learning is presented. First of all,a framework of non-negative matrix tri-factorization based on domain adaptive learning is constructed,in which the unchanged information is regarded as the bridge of knowledge transformation from the source domain to the target domain; The second step is to construct L1-Graph on the basis of sparse representation,adaptively search neighbor data and preserve the geometric structure of samples and features; Lastly,two complementary objective functions are integrated into a unified optimization problem,and then an iterative algorithm is adopted to solve the optimization problem,and the category of the test samples is estimated.Three comparative experiments were conducted on USPS-Binary handwritten digit dataset,three-Domain Object Benchmark dataset and ALOI dataset,the experiment results verify the effectiveness of the proposed algorithm,which improves the recognition accuracy and also ensures the robustness of the algorithm.