現行的深度學習分類模型是從一批標註的資料學習各類別的特徵後，在測試階段模型預測輸入資料屬於哪一種類別。由於模型往往依靠最終各類別的數值大小來決定其類別，一旦遇到輸入資料和原訓練資料的分布差異甚大時，模型也只能依照原訓練途徑產生差強人意的類別預測，這是我們不樂見的現象。領域適應任務旨在設計各種「適應」策略，以解決不同領域之間資料分布差異的遷移學習能力問題。
本研究聚焦於時序資料的非監督式領域適應任務，以向量量化變分自動編碼器為基底，並分成三種不同的階段進行模型的訓練、適應以及推論。在訓練階段，我們將源域資料經過雙層式編碼器分別萃取時域和頻域的潛在表徵，再透過初始化的階層式特徵向量表分配每個潛在表徵至對應的類別區間，並利用最近鄰居演算法搜尋最相似的特徵向量作為解碼器的輸入。在適應階段由於目標域沒有類別標籤，因此，對於每個潛在表徵的最近鄰居演算法搜尋範圍將擴展成完整的特徵向量表，再加入不同的損失函數作為限制以避免模型崩潰。在推論階段中，為了增加模型的穩定性，模型最終會以最近鄰居演算法搜尋的前五名特徵向量所對應的類別進行投票決定其類別。
本研究另一項重點在於如何調控時域與頻域對於不同資料分布的敏感度以非同步式分別更新階層式特徵向量表的時域與頻域區塊。我們利用頻域與時域的特性，頻域更能彰顯類別而非領域相關的特徵資訊，時域則能提供更多細節資訊，使模型更能夠學習分類。
本研究進行了八項實驗用以評估模型的性能、實用性以及潛力。在非監督領域適應的五種資料集實驗中，本研究的模型性能在分類準確性上高於現行最佳模型平均約十個百分點。而在另一組領域泛化任務的實驗中，雖然表現不如目前表現最優異的方法，然而位居第二的表現仍證明我們的模型是極具潛力的。接著透過消融實驗，得以讓我們更清楚知道架構中的每個模組相互作用對於模型的表現。最後，為了驗證本研究提出的架構能夠實際應用在真實場域中，本研究進行透析中低血壓的預測任務，其中訓練和測試數據來自不同醫療院所。在大部分的實驗中，本研究的模型架構表現皆優於前人研究。

Current deep-learning classification models learn the features of each category from a batch of labeled data, and during the testing phase, the model predicts which category the input data belongs to. Since the model often relies on the magnitude of the final values for each category to determine its classification, when the input data distribution differs significantly from the original training data, the model can only produce unsatisfactory predictions based on the original model weights. Domain adaptation tasks aim to design various "adaptation" strategies to solve the problem of transfer learning ability across data distributions from different domains.
This research focuses on unsupervised domain adaptation tasks for time-series data, using a vector-quantized variational autoencoder as the backbone, and divides the process into three different stages: training, adaptation, and inference. During the training phase, we extract latent representations of the temporal and frequency domains separately through a dual-stream encoder for the source domain data. These latent representations are then assigned to corresponding category intervals through an initialized hierarchical embedding table, and the nearest neighbor algorithm is used to find the most similar embeddings as input to the decoder. In the adaptation phase, since the target domain lacks class labels, the search range of the nearest neighbor algorithm for each latent representation will be expanded to the entire table, and different loss functions will be added as constraints to avoid model collapse. During the inference phase, to increase the model's stability, the model will ultimately use the top five embeddings found by the nearest neighbor algorithm to vote for the corresponding class.
Another key focus of this research is how to adjust the sensitivity of the temporal and frequency domains to different data distributions by asynchronously updating the temporal and frequency domain blocks of the hierarchical embedding table. We utilize the characteristics of the frequency and temporal domains—the frequency domain often highlights class-specific features rather than domain-variant features, while the temporal domain provides more detailed information, enabling the model to learn classifications more effectively.
We conducted eight experiments to evaluate the model’s performance, utility, and potential. In five unsupervised domain adaptation datasets, our model outperformed the current state-of-the-art models by an average of about 10%. While our model did not surpass the best-performing method in another set of domain generalization tasks, it achieved second place, demonstrating significant potential. Ablation experiments clarified the interaction and impact of each module within the framework on overall model performance. Finally, to validate the practical application of our proposed framework, we conducted a task predicting hypotension during dialysis using training and testing data from different institutions. In most experiments, our model outperformed previous studies.

[1] M. Ragab, E. Eldele, W. L. Tan, C. S. Foo, Z. Chen, M. Wu, and X. Li. Adatime: A benchmarking suite for domain adaptation on time series data, 2023.
[2] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017.
[3] Andrew S Levey, Lesley A Stevens, Christopher H Schmid, Yaping Zhang, Alejandro F Castro III, Harold I Feldman, John W Kusek, Paul Eggers, Frederick Van Lente, Tom Greene, et al. A new equation to estimate glomerular filtration rate. Annals of internal medicine, 150(9):604–612, 2009.
[4] H. He, O. Queen, T. Koker, C. Cuevas, T. Tsiligkaridis, and M. Zitnik. Domain adaptation for time series under feature and label shifts, 2023.
[5] Ruichu Cai, Jiawei Chen, Zijian Li, Wei Chen, Keli Zhang, Junjian Ye, Zhuozhang Li, Xiaoyan Yang, and Zhenjie Zhang. Time series domain adaptation via sparse associative structure alignment, 2021.
[6] Ngai Hang Chan. Time series: applications to finance. John Wiley & Sons, 2004.
[7] Ray Ball and Ross Watts. Some time series properties of accounting income. The Journal of Finance, 27(3):663–681, 1972.
[8] H Lehman Li-wei, Ryan P Adams, Louis Mayaud, George B Moody, Atul Malho tra, Roger G Mark, and Shamim Nemati. A physiological time series dynamics-based approach to patient monitoring and outcome prediction. IEEE journal of biomedical and health informatics, 19(3):1068–1076, 2014.
[9] Vasileios Lampos, Maimuna S Majumder, Elad Yom-Tov, Michael Edelstein, Si mon Moura, Yohhei Hamada, Molebogeng X Rangaka, Rachel A McKendry, and Ingemar J Cox. Tracking covid-19 using online search. NPJ digital medicine, 4(1):17, 2021.
[10] Olufemi Olominu and Abiodun A Sulaimon. Application of time series analysis to predict reservoir production performance. In SPE Nigeria Annual International Conference and Exhibition, pages SPE–172395. Society of Petroleum Engineers, 2014.
[11] Hadi Hajmohammadi and Benjamin Heydecker. Multivariate time series modelling for urban air quality. Urban Climate, 37:100834, 2021.
[12] Lisa Torrey and Jude Shavlik. Transfer learning. In Handbook of Research on Ma chine Learning Applications and Trends: Algorithms, Methods, and Techniques, pages 242–264. IGI Global, 2010.
[13] Jose G. Moreno-Torres, Teresa Raeder, Rosana Alaiz-Rodr´ıguez, Nitesh V. Chawla, and Francisco Herrera. A unifying view on dataset shift in classifica tion. Pattern Recognition, 45:521–530, 2012.
[14] Uday Kamath, Jingjing Liu, and John Whitaker. Transfer learning: Domain adaptation. In Deep Learning for NLP and Speech Recognition, pages 495–535. Springer, 2019.
[15] Eric Tzeng, Judy Hoffman, Ning Zhang, Kate Saenko, and Trevor Darrell. Deep domain confusion: Maximizing for domain invariance, 2014.
[16] Baochen Sun and Kate Saenko. Deep coral: Correlation alignment for deep domain adaptation, 2016.
[17] B. B. Damodaran, B. Kellenberger, R. Flamary, D. Tuia, and N. Courty. Deepjdot: Deep joint distribution optimal transport for unsupervised domain adaptation, 2018.
[18] Chao Chen, Zhihang Fu, Zhihong Chen, Sheng Jin, Zhaowei Cheng, Xinyu Jin, and Xian-Sheng Hua. Homm: Higher-order moment matching for unsupervised domain adaptation, 2020.
[19] Mohammad Mahfujur Rahman, Clinton Fookes, Mahsa Baktashmotlagh, and Sridha Sridharan. On minimum discrepancy estimation for deep domain adap tation. In Domain Adaptation for Visual Understanding, pages 81–94. 2020.
[20] Y. Ganin, E. Ustinova, H. Ajakan, P. Germain, H. Larochelle, F. Laviolette, et al. Domain-adversarial training of neural networks, 2016.
[21] Mingsheng Long, Zhangjie Cao, Jianmin Wang, and Michael I Jordan. Conditional adversarial domain adaptation, 2018.
[22] Rui Shu, Hung H Bui, Hirokazu Narui, and Stefano Ermon. A dirt-t approach to unsupervised domain adaptation, 2018.
[23] Guoliang Kang, Lu Jiang, Yi Yang, and Alexander G Hauptmann. Contrastive adaptation network for unsupervised domain adaptation, 2019.
[24] Ankit Singh. Clda: Contrastive learning for semi-supervised domain adaptation, 2021.
[25] Y. Zhang, Z. Wang, J. Li, J. Zhuang, and Z. Lin. Towards effective instance discrimination: Contrastive loss for unsupervised domain adaptation, 2023.
[26] S. Purushotham, Wilka Carvalho, Tanachat Nilanon, and Yan Liu. Variational recurrent adversarial deep domain adaptation, 2017.
[27] Garrett Wilson, Janardhan Rao Doppa, and Diane Joyce Cook. Multi-source deep domain adaptation with weak supervision for time-series sensor data, 2020.
[28] Qiao Liu and Hui Xue. Adversarial spectral kernel matching for unsupervised time series domain adaptation, 8 2021.
[29] Felix Ott, David Rugamer, Lucas Heublein, Bernd Bischl, and Christopher Mutschler. Domain adaptation for time series classification to mitigate covari ate shift, 2022.
[30] Xiaoyong Jin, Youngsuk Park, Danielle Maddix, Hao Wang, and Yuyang Wang. Domain adaptation for time series forecasting via attention sharing, 17–23 Jul 2022.
[31] Yilmazcan Ozyurt, Stefan Feuerriegel, and Ce Zhang. Contrastive learning for unsupervised domain adaptation of time series, 2022.
[32] E. Eldele, M. Ragab, Z. Chen, M. Wu, C. K. Kwoh, and X. Li. Contrastive domain adaptation for time-series via temporal mixup, 2023.
[33] G. Wilson, J. R. Doppa, and D. J. Cook. Calda: Improving multi-source time series domain adaptation with contrastive adversarial learning, 2023.
[34] A. Van Den Oord and O. Vinyals. Neural discrete representation learning, 2017.
[35] Davide Anguita, Alessandro Ghio, Luca Oneto, Xavier Parra, and Jorge Luis Reyes-Ortiz. A public domain dataset for human activity recognition using smart phones, 2013.
[36] Allan Stisen, Henrik Blunck, Sourav Bhattacharya, Thor Siiger Prentow, Mikkel Baun Kjærgaard, Anind Dey, Tobias Sonne, and Mads Møller Jensen. Smart devices are different: Assessing and mitigating mobile sensing hetero geneities for activity recognition, 2015.
[37] Jennifer R. Kwapisz, Gary M. Weiss, and Samuel A. Moore. Activity recognition using cell phone accelerometers, 2011.
[38] Ary L. Goldberger, Luis A. N. Amaral, Leon Glass, Jeffrey M. Hausdorff, Pla men Ch. Ivanov, Roger G. Mark, Joseph E. Mietus, George B. Moody, Chung Kang Peng, and H. Eugene Stanley. Physiobank, physiotoolkit, and physionet: Components of a new research resource for complex physiologic signals, 2000.
[39] Christian Lessmeier, James Kuria Kimotho, Detmar Zimmer, and Walter Sextro. Condition monitoring of bearing damage in electromechanical drive systems by using motor current signals of electric motors: A benchmark data set for data driven classification, 2016.
[40] X. Qin, J. Wang, S. Ma, W. Lu, Y. Zhu, X. Xie, and Y. Chen. Generalizable low-resource activity recognition with diverse and discriminative representation learning, 2023.
[41] W. Lu, J. Wang, X. Sun, Y. Chen, and X. Xie. Out-of-distribution representation learning for time series classification, September 2022.
[42] Billur Barshan and Murat Cihan Yuksek. Recognizing daily and sports activities in two open source machine learning environments using body-worn sensor units, 2014.
[43] Vladimir Vapnik. Principles of risk minimization for learning theory, 1992.
[44] Minghao Xu, Jian Zhang, Bingbing Ni, Teng Li, Chengjie Wang, Qi Tian, and Wenjun Zhang. Adversarial domain adaptation with domain mixup, 2020.
[45] Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy M Hospedales. Learning to generalize: Meta-learning for domain generalization, 2018.
[46] Zeyi Huang, Haohan Wang, Eric P Xing, and Dong Huang. Self-challenging im proves cross-domain generalization, 2020.
[47] Giambattista Parascandolo, Alexander Neitz, Antonio Orvieto, Luigi Gresele, and Bernhard Sch¨olkopf. Learning explanations that are hard to vary, 2021.
[48] Yuge Shi, Jeffrey Seely, Philip HS Torr, N Siddharth, Awni Hannun, Nicolas Usunier, and Gabriel Synnaeve. Gradient matching for domain generalization, 2022.
[49] 應伶軒. 應用時間感知注意力網路進行透析中低血壓的個人化預測. 國立成功大 學碩士論文, 2023.