基于张量分解的域适应算法

Domain Adaptation Algorithm Based on Tensor Decomposition

由于训练数据易过期,在多数情况下训练数据和测试数据具有不同的特征分布,因此在利用源域信息时,须先尽量减小不同领域的特征分布的差异。使用张量表示特征可以维持高维空间数据的本征结构信息。朴素张量子空间学习法虽然是面向张量特征的域适应方法,但其复杂度较高,且没有达到较好的知识迁移效果。为此,文中提出了基于张量分解的域适应算法,即张量列子空间学习法和张量环子空间学习法,二者的主要思想相似。首先,使用张量表示源域和目标域的特征;其次利用张量分解方法,将特征分解为一系列三阶张量来表示子空间;然后,依次将源域特征和目标域特征映射到子空间中;最后,将特征张量重塑为矩阵形式,基于映射后的源域特征训练模型,基于映射后的目标域特征完成新领域的任务。实验结果表明,在无监督图像分类中,张量列子空间学习法和张量环子空间学习法在准确率和运行时间方面都有所提升。相比于朴素张量子空间学习法,张量列子空间学习法和张量环子空间学习法的准确率分别提高了1.68%和2.08%,且运行时间也有明显减少,算法复杂度较小。实验数据充分说明,基于张量分解的域适应算法充分减小了源域特征和目标域特征之间的差异,实现了不同领域间的知识复用。

Because training data tend to be outdated,training data and test data have different feature distributions in most cases.Therefore,when using the source domain information,it is necessary to minimize the difference of feature distributions in different fields.Features represented by tensor can maintain the intrinsic structure information of high-dimensional spatial data.Naive tensor subspace learning is a domain adaptation method for tensor features,but it has high complexity and can not achieve good knowledge transfer effect.For this reason,this paper proposed a domain adaptation algorithm based on tensor decomposition,namely tensor train subspace learning and tensor ring subspace lear-ning,and the main ideas of the two methods are similar.Firstly,the features of source domain and target domain are coded into tensor.By using the tensor decomposition,the tensor of features is decomposed into a series of third-order tensors to represent the subspace.Then,the features of source domain and target domain are mapped into subspace successively.Finally,the feature tensor is reshaped into matrix form.Based on the mapped features of source domain training model and the mapped feature of target domain,the task in new domain is completed.Experiments show that the tensor train subspace learning and the tensor ring subspace learning are improved in terms of accuracy and running time for unsupervised image classification.Compared with the naive tensor subspace learning,the accuracy of the tensor train subspace learning and the tensor ring subspace learning is improved by 1.68%and 2.08%respectively,the running time is also reduced significantly,and the complexity of the algorithm is smaller.Experimental results show that the domain adaptation algorithm based on tensor decomposition can reduce the difference between source domain and target domain,and realize the reuse of knowledge between different domains.