| 摘要: |
| 近年来深度学习因其在各个场景下的优异性能而受到越来越多研究者的重视, 但是这些方法通常依赖独立同分布假设. 领域自适应则是为了缓解分布偏移带来的影响而提出的问题, 它利用带标签的源域数据和不带标签的目标域数据能够训练得到在目标域数据上性能较好的模型. 现有的领域自适应方法大多针对静态数据, 而时间序列数据的方法需要捕捉变量之间的依赖关系. 现有的方法虽然采用针对时间序列数据的特征提取器, 例如递归神经网络, 以学习变量间的依赖关系, 但是往往将冗余的信息也一同提取. 这些冗余信息容易和语义信息耦合, 进而影响模型的预测性能. 基于上述问题, 提出一种基于路径签名的时间序列领域自适应方法(path-signature-based time-series domain adaptation, PSDA). 该方法一方面利用路径签名变换来捕捉变量间的稀疏依赖关系, 排除冗余相关关系的同时保留语义相关关系, 从而有利于提取时序数据中具有判别力的特征; 另一方面通过约束源域和目标域之间的依赖关系一致性来保留领域之间不变的依赖关系, 排除领域变化的依赖关系, 从而有利于提取时序数据中具有泛化性的特征. 基于以上方法, 进一步提出一个距离度量标准和泛化性边界理论, 并且在多个时间序列领域自适应标准数据集上获得了最好的实验效果. |
| 关键词: 迁移学习 时间序列领域自适应 路径签名 |
| DOI：10.13328/j.cnki.jos.007149 |
| 分类号:TP18 |
| 基金项目:科技创新2030—“新一代人工智能”重大项目(2021ZD0111501); 国家优秀青年科学基金 (62122022); 国家自然科学基金(61876043, 61976052, 62206064) |
|
| Time-series Domain Adaptation Based on Path Signature |
|
CAI Rui-Chu1, YAN Jia-Wen1, CHEN Dao-Xin1, LI Zi-Jian1, HAO Zhi-Feng2
|
|
1.School of Computer Science and Technology, Guangdong University of Technology, Guangzhou 510006, China;2.College of Engineering, Shantou University, Shantou 515063, China
|
| Abstract: |
| Recently, deep learning has received increasing attention from researchers due to its excellent performance in various scenarios, but these methods often rely on the independent and identically distribution assumption. Domain adaptation is a problem proposed to mitigate the impact of distribution offset, which uses labeled source domain data and unlabeled target domain data to achieve better performance on target data. Existing methods are devised for static data, while the methods for time series data need to capture the dependencies between variables. Although these methods use feature extractors for time series data, such as recurrent neural networks, to learn the dependencies between variables, they often extract redundant information. This information is easily entangled with semantic information, affecting the model performance. To solve these problems, this study proposes a path-signature-based time-series domain adaptation (PSDA). On the one hand, this method uses path signature transformation to capture sparse dependencies between variables and eliminate redundant correlations while preserving semantic dependencies, thereby facilitating the extraction of discriminative features from temporal data. On the other hand, the invariant dependency relationships are preserved by constraining the consistency of dependency relationships among different domains, and the changing dependency relationships between domains are excluded, which is conducive to extracting generalized features from temporal data. Based on the above methods, the study further proposes a distance metric and generalized boundary theory and obtains the best experimental results on multiple time series domain adaptation standard datasets. |
| Key words: transfer learning time-series domain adaptation path signature |
