Abstract:Domain generalization aims to design models that can effectively generalize to unseen target domains by learning from observed source domains. Domain generalization poses a significant challenge for time series data, due to varying data distributions and temporal dependencies. Existing approaches to domain generalization are not designed for time series data, which often results in suboptimal or unstable performance when confronted with diverse temporal patterns and complex data characteristics. We propose a novel approach to tackle the problem of domain generalization in time series forecasting. We focus on a scenario where time series domains share certain common attributes and exhibit no abrupt distribution shifts. Our method revolves around the incorporation of a key regularization term into an existing time series forecasting model: domain discrepancy regularization . In this way, we aim to enforce consistent performance across different domains that exhibit distinct patterns. We calibrate the regularization term by investigating the performance within individual domains and propose the domain discrepancy regularization with domain difficulty awareness . We demonstrate the effectiveness of our method on multiple datasets, including synthetic and real-world time series datasets from diverse domains such as retail, transportation, and finance. Our method is compared against traditional methods, deep learning models, and domain generalization approaches to provide comprehensive insights into its performance. In these experiments, our method showcases superior performance, surpassing both the base model and competing domain generalization models across all datasets. Furthermore, our method is highly general and can be applied to various time series models.

What problem does this paper attempt to address?

### Problems the Paper Attempts to Solve This paper aims to address the issue of Domain Generalization in time series forecasting. Specifically, the authors focus on how to design models that can effectively generalize to unseen target domains, constructed by learning from observed source domains. Time series data, due to its variability in distribution and temporal dependencies, presents significant challenges for domain generalization. Existing domain generalization methods are often not suitable for time series data, leading to poor or unstable performance when faced with diverse temporal patterns and complex data characteristics. To tackle this challenge, the authors propose a novel approach to handle the domain generalization problem in time series forecasting. They focus on a scenario where there are certain common attributes between different time series domains and no abrupt distribution changes. The core of this method is the introduction of a key regularization term in existing time series forecasting models: Domain Discrepancy Regularization. Through this approach, the authors aim to achieve consistent performance across different domains. Additionally, they propose an extended version of the regularization term, Domain Discrepancy Regularization with Domain Difficulty Awareness, to further enhance the model's generalization capability. ### Main Contributions 1. **Proposing a New Domain Generalization Problem**: The authors define a new domain generalization problem in the context of time series forecasting and formalize the time series forecasting task under specific assumptions, laying the foundation for subsequent research. 2. **Introducing Domain Discrepancy Regularization**: The authors propose a new regularization term to improve the generalization capability of forecasting models by adjusting cross-domain performance discrepancies. 3. **Introducing Domain Discrepancy Regularization with Domain Difficulty Awareness**: The authors further extend the regularization term by considering the difficulty of each domain, allowing the model to learn more complex patterns in complex domains. ### Experimental Validation The authors conducted extensive experiments on multiple synthetic and real-world time series datasets, including datasets from retail, transportation, and finance domains. The experimental results demonstrate that their method exhibits higher accuracy in domain generalization tasks and has lower training overhead, making it suitable for real-world scenarios.

Domain Generalization in Time Series Forecasting