DualDynamics: Synergizing Implicit and Explicit Methods for Robust Irregular Time Series Analysis

YongKyung Oh, Dong‐Young Lim, Sungil Kim

Abstract

Real-world time series analysis faces significant challenges when dealing with irregular and incomplete data. While Neural Differential Equation (NDE) based methods have shown promise, they struggle with limited expressiveness, scalability issues, and stability concerns. Conversely, Neural Flows offer stability but falter with irregular data. We introduce 'DualDynamics', a novel framework that synergistically combines NDE-based method and Neural Flow-based method. This approach enhances expressive power while balancing computational demands, addressing critical limitations of existing techniques. We demonstrate DualDynamics' effectiveness across diverse tasks: classification of robustness to dataset shift, irregularly-sampled series analysis, interpolation of missing data, and forecasting with partial observations. Our results show consistent outperformance over state-of-the-art methods, indicating DualDynamics' potential to advance irregular time series analysis significantly.

The Representation Gap in Time Series Analysis

Real-world time series—such as medical records or financial tickers—rarely arrive in neat, equally spaced intervals. They are often messy, irregular, and characterized by significant missingness. Current approaches typically fall into two categories:

  1. Implicit Methods (Neural Differential Equations): These treat data as a continuous process, excelling at handling irregular gaps. However, they are often computationally expensive ("stiff") and struggle to represent complex, high-frequency patterns.
  2. Explicit Methods (Neural Flows): These map initial states to future states directly via neural networks. While fast and stable, they generally assume regular sampling and struggle with noisy or unknown initial states.

DualDynamics addresses this "representation gap" by creating a synergistic architecture that leverages the strengths of both paradigms.

Methods: A Two-Stage Latent Representation

The core innovation of DualDynamics is a two-stage latent representation optimized jointly via adjoint-based backpropagation.

DualDynamics Architecture: Synergizing Implicit (NDE) and Explicit (Neural Flow) components.
  • The Implicit Component (The "Bridge"): The framework utilizes a Neural Controlled Differential Equation (Neural CDE) to map irregular raw input into a continuous primary latent space ($z$). This component handles the mathematical complexity of interpolating between observations.
  • The Explicit Component (The "Refiner"): The primary latent state $z$ is then fed into a Neural Flow (e.g., ResNet Flow or GRU Flow). This component acts as a diffeomorphism—a smooth, invertible transformation—reshaping the latent space into a more expressive secondary latent space ($\hat{z}$).
  • Joint Optimization: DualDynamics optimizes both components simultaneously. This allows the NDE to learn how to best prepare data for the Flow, while the Flow learns to refine the NDE's output for the specific task.

Key Results and Performance

The framework was evaluated across a "decathlon" of benchmarks, including 18 UEA/UCR datasets, PhysioNet clinical data, MuJoCo physics simulations, and Google Stock data.

Performance Comparison: DualDynamics maintains high accuracy across varying missing data rates compared to baselines.
  • Classification: On 18 diverse datasets, DualDynamics achieved a top-ranking average accuracy of 0.708, outperforming Neural CDE (0.667) and specialized attention models like MTAN (0.640). As shown in Figure 2, it maintains superior performance even as the missingness rate increases to 70%.
  • Interpolation: On the PhysioNet Mortality task, it achieved a Mean Squared Error (MSE) of 3.114 at 90% missingness, significantly lower than the previous state-of-the-art mTAND-Full (4.798).
  • Forecasting: In the MuJoCo environment, the model reduced forecasting error by nearly 70% compared to standard Neural CDEs.

Insights and Discussion

One of the most striking findings is the stability of the training process. The addition of the Flow component acts as a regularizer, leading to smoother convergence compared to standard NDEs, which often exhibit "spiky" loss curves. Furthermore, the performance gain is structural; DualDynamics outperforms standard MLPs even when parameter counts are matched, highlighting the value of the flow's invertibility and smoothness.

While DualDynamics represents a significant advancement, selection of the optimal flow type currently requires hyperparameter tuning. Additionally, while more efficient than deep NDE stacks, the reliance on ODE solvers remains a potential bottleneck for extremely long sequences. Future work may explore replacing the explicit component with Transformer-based mechanisms to further push the boundaries of expressiveness.

Summary updated

Abstract

Real-world time series analysis faces significant challenges when dealing with irregular and incomplete data. While Neural Differential Equation (NDE) based methods have shown promise, they struggle with limited expressiveness, scalability issues, and stability concerns. Conversely, Neural Flows offer stability but falter with irregular data. We introduce 'DualDynamics', a novel framework that synergistically combines NDE-based method and Neural Flow-based method. This approach enhances expressive power while balancing computational demands, addressing critical limitations of existing techniques. We demonstrate DualDynamics' effectiveness across diverse tasks: classification of robustness to dataset shift, irregularly-sampled series analysis, interpolation of missing data, and forecasting with partial observations. Our results show consistent outperformance over state-of-the-art methods, indicating DualDynamics' potential to advance irregular time series analysis significantly.
BibTeX
@inproceedings{oh2025dualdynamics,
  author = {YongKyung Oh and Dong‐Young Lim and Sungil Kim},
  title = {{DualDynamics: Synergizing Implicit and Explicit Methods for Robust Irregular Time Series Analysis}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.1609/aaai.v39i18.34173},
}
Figure