Aligning Time-series by Local Trends: Applications in Public Health

Ajitesh SrivastavaAAAI 2025 · DOI

TL;DR

This paper introduces DTW+SBA, a novel method for creating ensemble time-series models that better capture the timing and severity of events, particularly crucial for public health communication. By representing time-series as sequences of local trends and aligning these trends, DTW+SBA overcomes the limitations of traditional mean ensembles that often misrepresent critical features like epidemic peaks.

What Problem Does This Paper Address?

In public health, forecasting epidemic trajectories often involves combining predictions from multiple models. This is typically done through "ensembling," where a single consensus trajectory is created. However, standard ensembling methods, like taking the mean of all trajectories at each time point, can obscure important details. For instance, if multiple models predict similar epidemic peaks but at slightly different times, a mean ensemble might produce a trajectory with a significantly lower and wider peak, misrepresenting the true severity and timing of the epidemic. This can lead to misjudgments in policy-making and public communication. The paper addresses the need for an ensembling method that preserves the key characteristics of individual trajectories, especially the timing and magnitude of significant events like epidemic peaks.

Methods

The core of the proposed method, DTW+SBA, lies in a novel way of representing and aligning time-series. Instead of directly comparing raw time-series values, the authors first transform each time-series into a "Shapelet-space Representation" (SSR). This SSR captures local trends within the time-series by measuring its similarity to a predefined set of "shapelets" (e.g., "increase," "peak," "surge," "flat"). This representation is generated by sliding a window across the time-series and computing the SSR for each window.

The alignment of these SSRs is then performed using Dynamic Time Warping (DTW), but with a crucial modification: DTW is applied to the SSRs rather than the raw time-series. This "DTW+S" alignment focuses on matching similar local trends, regardless of their exact timing. Finally, these aligned local trends are used to construct an ensemble using a barycenter averaging approach (DTW+SBA). This process effectively averages the timing and severity of corresponding local trends across all input trajectories, leading to a more representative ensemble. The authors also provide theoretical guarantees on the properties of their shapelet-space representation, ensuring it preserves "closeness" (similar trends are mapped to nearby points).

Datasets

The paper evaluates its method on 12 sets of real-world time-series data. These datasets are derived from influenza and RSV modeling efforts, specifically from:

  • Influenza Scenario Modeling Hubs: Including data from US SMH 2020, European CDC 2022, and US SMH 2022.
  • RSV Scenario Modeling Hubs: Data from RSV Scenario Modeling Hub Round 1.

These datasets consist of multiple trajectories (ranging from 75 to over 1000 per set) generated by different models under various scenarios, making them suitable for evaluating ensembling and clustering techniques.

Key Insights and Ideas

  • Local Trends Matter More Than Raw Values for Ensembling: The paper convincingly argues that focusing on the shape and timing of local trends, rather than just the average value at a specific time point, is critical for creating meaningful ensembles, especially in dynamic processes like epidemics.
  • Shapelet-Space Representation Enables Interpretable Alignment: By transforming time-series into a shapelet-space, the method achieves an interpretable representation where each dimension relates to a specific trend. This allows for alignment based on these trends, making the ensembling process more intuitive.
  • DTW+SBA Preserves Peak Dynamics: The most striking insight is how DTW+SBA effectively captures and preserves the timing and magnitude of peaks in ensemble trajectories, a significant improvement over traditional mean ensembles that tend to flatten them.
  • Clustering Benefits from Trend-Aware Similarity: The paper demonstrates that using DTW+S as a distance metric for clustering time-series leads to more coherent and meaningful clusters, separating trajectories based on their underlying trend patterns rather than just overall similarity.

Key Results

  • Improved Peak Representation: Figure 4(a) visually shows that the DTW+SBA ensemble produces a significantly higher peak compared to the mean ensemble and DTW (z-norm) BA, which tend to flatten the peak.
  • Reduced Fractional Error: Table 1 quantifies this improvement, showing that DTW+SBA achieves the lowest fractional error in estimating both peak size and timing across various datasets, indicating it better captures the ground truth of individual trajectories.
  • Better Clustering: Figure 6 demonstrates that DTW+S leads to more reasonable clustering of time-series data, effectively separating trajectories with similar trend patterns, whereas other methods (like normalized DTW or shapelet-only) mix different patterns into single clusters.

Limitations and Future Work

  • Domain Knowledge for Shapelets: The authors acknowledge that choosing appropriate shapelets might require domain expertise. However, they provide theoretical guidelines (Theorems 1 and 2) to ensure that chosen shapelets satisfy the desired "closeness-preserving" property.
  • Computational Complexity: The current implementation of DTW on the SSR has a time and space complexity of O(dwT), where T is time-series length, w is window size, and d is the number of shapelets.
  • Future Work: The authors suggest exploring optimizations for DTW (e.g., Keogh and Pazzani 2000) and adapting them to the DTW+S framework to improve computational efficiency. They also mention the potential for extending this approach to other fields where understanding local trends in time-series is important.

Summary updated

BibTeX
@inproceedings{srivastava2025aligning,
  author = {Ajitesh Srivastava},
  title = {{Aligning Time-series by Local Trends: Applications in Public Health}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.1609/aaai.v39i27.35062},
}
Figure