Abstract
Time series data is ubiquitous across various domains such as finance, healthcare, and manufacturing, but their properties can vary significantly depending on the domain they originate from. The ability to perform Content-based Time Series Retrieval (CTSR) is crucial for identifying unknown time series examples. However, existing CTSR works typically focus on retrieving time series from a single domain database, which can be inadequate if the user does not know the source of the query time series. This limitation motivates us to investigate the CTSR problem in a scenario where the database contains time series from multiple domains. To facilitate this investigation, we introduce a CTSR benchmark dataset that comprises time series data from a variety of domains, such as motion, power demand, and traffic. This dataset is sourced from a publicly available time series classification dataset archive, making it easily accessible to researchers in the field. We compare several popular methods for modeling and retrieving time series data using this benchmark dataset. Additionally, we propose a novel distance learning model that outperforms the existing methods. Overall, our study highlights the importance of addressing the CTSR problem across multiple domains and provides a useful benchmark dataset for future research.
Introduction: The Multi-Domain Retrieval Challenge
Imagine discovering a sensor recording without context—is it a power signature from a kitchen appliance, a heart rate monitor, or industrial machinery? Identifying such signals requires a Content-based Time Series Retrieval (CTSR) system capable of searching a massive, labeled database for the closest match.
Traditional CTSR research often assumes the query and database share the same domain (e.g., ECG-to-ECG). However, real-world scenarios frequently involve "unknown" signals. Time series from different domains exhibit vastly different characteristics, making multi-domain retrieval—distinguishing between motion, power, and traffic data—a significant challenge. This paper addresses the lack of robust systems and benchmarks for this multi-domain context.
Methods and the RN2D Architecture
The authors evaluated six baseline methods: traditional non-learning approaches like Euclidean Distance (ED) and Dynamic Time Warping (DTW), and deep learning models including LSTM, GRU, Transformers, and 1D Residual Networks (RN1D). The neural baselines typically utilize a "Siamese" architecture, encoding two time series independently to calculate similarity in a latent space.
To bridge the gap between alignment-aware methods (like DTW) and high-capacity neural networks, the authors propose RN2D (Residual Network 2D).
Unlike 1D sequence processors, RN2D first calculates a pairwise distance matrix between the query and the target, creating a 2D "map" of temporal relationships. This matrix is then treated as an image and processed through a 2D Residual Network. This approach allows the model to learn sophisticated distance functions by "seeing" complex alignment patterns, effectively combining the strengths of alignment-aware formulas and deep learning.
Multi-Domain Benchmark Dataset
A new CTSR benchmark was created by repurposing the UCR Archive, a collection of 128 time series classification datasets. The resulting benchmark is both large-scale and diverse:
- Scale: 136,377 training, 17,005 test, and 17,005 validation samples.
- Diversity: Spans domains like motion, power demand, and traffic, providing a rigorous test for cross-domain generalization.
Key Insights and Results
The evaluation yielded several critical findings:
- Alignment is Essential: DTW significantly outperformed Euclidean Distance, confirming that temporal shifts and "warping" are vital considerations even in multi-domain scenarios.
- Neural Scaling: All neural network baselines outperformed DTW, suggesting that learned features are more effective than raw data points for diverse data.
- The Hybrid Advantage: RN2D's use of 2D alignment matrices within a ResNet architecture captures the "best of both worlds," outperforming all other models.
- Model Selection: Interestingly, LSTMs slightly outperformed Transformers in this retrieval task, indicating that newer architectures are not always superior for every sequential problem.
Quantitatively, RN2D achieved state-of-the-art results across all primary metrics:
- NDCG@10: 0.9325
- Precision@10: 0.9266
- Mean Average Precision (AP@10): 0.9342
Qualitative analysis further showed that while simpler methods struggled with complex or periodic signals, RN2D reliably retrieved relevant items even amidst visual noise or temporal shifts.
Conclusion and Future Directions
While RN2D shows great promise, two main areas for future work remain:
- Efficiency: Processing 2D matrices is computationally intensive. Future research will explore hashing or acceleration techniques for large-scale deployment.
- Pretraining: The authors suggest that self-supervised pretraining (Foundation Models for time series) could further enhance retrieval accuracy beyond training from scratch.
Summary updated