Analyzing and Improving Diffusion Models for Time-Series Data Imputation: A Proximal Recursion Perspective

Zhichao Chen, Hao Wang, Fangyikang Wang, Licheng Pan, Zhengnan Li, Yunfei Teng, Haoxuan Li, Zhouchen LinarXiv 2026 · arXiv · OpenReview

TL;DR

This paper tackles the inconsistency of diffusion models (DMs) for time-series data imputation (TSDI) by reframing the problem through a "proximal recursion" lens. They introduce SPIRIT, a novel framework that replaces the standard Wasserstein distance with a more robust "semi-proximal transport" discrepancy and removes a dissipative term, leading to significantly improved imputation accuracy, especially in complex, non-stationary scenarios.

What Problem Does This Paper Address?

Time-series data imputation (TSDI) is crucial across many domains, from healthcare to industrial monitoring, where missing data can compromise analysis. Diffusion models (DMs) have shown promise for this task due to their generative capabilities. However, their performance in real-world, complex time-series data remains inconsistent. The authors identify two key issues:

  1. Non-stationarity: Time-series data often exhibit changing statistical properties over time. Standard DM formulations, particularly their reliance on Wasserstein distance, can be sensitive to these shifts, leading to biased imputation trajectories and poor handling of outliers.
  2. Objective Inconsistency: DMs are inherently designed for generating diverse samples, which can conflict with the TSDI goal of accurately recovering missing values (pointwise fidelity). This can lead to imputations that are too diverse and less accurate.

The paper aims to understand these limitations from a theoretical perspective and propose a more robust and accurate DM-based imputation method.

Methods

The core of the paper lies in analyzing DM-based TSDI through the lens of proximal recursion. This framework views optimization problems as iterative applications of a "proximal operator." The authors identify two problematic components in existing DM-based TSDI:

  • Wasserstein Distance Regularization: This term, inherent in many DM formulations, enforces strict mass conservation, making it brittle to non-stationary data and outliers.
  • Dissipative Structure: This component, related to the SDE dynamics of DMs, promotes sample diversity but can reduce imputation accuracy.

To address these, the authors propose:

  1. Semi-Proximal Transport (SPT) Discrepancy: They relax the strict mass-preserving constraint of the Wasserstein distance by introducing an entropy-induced generalized Bregman divergence. This allows for more flexible "transport plans" that are theoretically proven to be robust against non-stationarity and outliers.
  2. Dissipative Structure Elimination: They remove the problematic dissipative term from the DM objective.

These two modifications form the basis of their proposed framework, SPIRIT (Semi-Proximal Transport Regularized time-series Imputation). SPIRIT reformulates TSDI as an optimization problem solved iteratively using the SPT discrepancy as the proximal operator.

Datasets

The authors evaluate their method on several standard public benchmarks for time-series imputation:

  • ETT dataset: (four subsets: ETTh1, ETTh2, ETTm1, ETTm2) - related to electricity transformer data.
  • Exchange: Daily exchange rates.
  • Illness: Weekly influenza-like illness data.
  • Traffic: Hourly occupancy rates from San Francisco Bay Area freeways.

Key Insights and Ideas

  • Proximal Recursion as a Unifying Lens: Viewing DM-based TSDI through proximal recursion reveals specific problematic terms (Wasserstein distance and dissipative structure) that hinder performance.
  • Wasserstein Distance's Achilles' Heel: The strict mass conservation in standard Wasserstein distance makes it ill-suited for non-stationary time-series data, leading to misalignments and sensitivity to outliers.
  • Diversity vs. Fidelity Trade-off: The inherent diversity-promoting nature of DMs can be detrimental to the accuracy-focused goal of imputation. Removing the dissipative structure is key to improving fidelity.
  • SPT for Robustness: The proposed Semi-Proximal Transport discrepancy offers a theoretically sound way to handle non-stationarity and outliers by relaxing mass conservation constraints.
  • SPIRIT's Dual Contribution: The framework's success stems from both the improved SPT discrepancy and the removal of the dissipative term, highlighting their complementary roles.

Key Results

  • State-of-the-Art Performance: SPIRIT achieves the best or second-best performance in 10 out of 14 cases across various datasets and missing ratios, demonstrating significant improvements over existing methods.
  • Robustness to Non-stationarity: Ablation studies show that removing the dissipative structure and introducing SPT are crucial for performance, especially on datasets prone to outliers.
  • Efficiency: The paper reports that SPIRIT's overall runtime remains below 3 minutes even for large datasets, with the "Score Learning" stage being the dominant factor.
  • Theoretical Guarantees: The paper provides theoretical analysis for the robustness of SPT and the convergence properties of the SPIRIT framework.

Limitations and Future Work

  • Periodic Structure and Temporal Dependence: The current approach treats time-series data as multivariate sequences without explicitly modeling temporal autocorrelation or periodic patterns. Future work could incorporate frequency-domain diffusion models or design proximal terms that better preserve periodicity.
  • Limited Uncertainty Quantification: Prioritizing accurate imputation by reducing the dissipative component might diminish sample diversity and weaken uncertainty estimates. Multi-objective optimization could be explored to balance accuracy and diversity for better uncertainty quantification.
  • Downstream-Task Awareness: The current focus is solely on TSDI. Integrating SPIRIT with downstream models (e.g., forecasting) through joint training or end-to-end objectives could enhance robustness in missing-data scenarios.
  • Beyond MCAR: The current formulation is for Missing Completely At Random (MCAR). Extending SPIRIT to Missing Not At Random (MNAR) settings would require modeling the missingness mechanism and incorporating distributionally robust optimization.
  • Alternative Score-Learning Strategies: The current use of denoising score matching incurs significant backpropagation overhead. Exploring alternative score-learning objectives or architectures could reduce computational costs.

Summary updated

Abstract

Diffusion models (DMs) have shown promise for Time-Series Data Imputation (TSDI); however, their performance remains inconsistent in complex scenarios. We attribute this to two primary obstacles: (1) non-stationary temporal dynamics, which can bias the inference trajectory and lead to outlier-sensitive imputations; and (2) objective inconsistency, since imputation favors accurate pointwise recovery whereas DMs are inherently trained to generate diverse samples. To better understand these issues, we analyze DM-based TSDI process through a proximal-operator perspective and uncover that an implicit Wasserstein distance regularization inherent in the process hinders the model's ability to counteract non-stationarity and dissipative regularizer, thereby amplifying diversity at the expense of fidelity. Building on this insight, we propose a novel framework called SPIRIT (Semi-Proximal Transport Regularized time-series Imputation). Specifically, we introduce entropy-induced Bregman divergence to relax the mass preserving constraint in the Wasserstein distance, formulate the semi-proximal transport (SPT) discrepancy, and theoretically prove the robustness of SPT against non-stationarity. Subsequently, we remove the dissipative structure and derive the complete SPIRIT workflow, with SPT serving as the proximal operator. Extensive experiments demonstrate the effectiveness of the proposed SPIRIT approach.
BibTeX
@misc{chen2026analyzing,
  author = {Zhichao Chen and Hao Wang and Fangyikang Wang and Licheng Pan and Zhengnan Li and Yunfei Teng and Haoxuan Li and Zhouchen Lin},
  title = {{Analyzing and Improving Diffusion Models for Time-Series Data Imputation: A Proximal Recursion Perspective}},
  year = {2026},
  eprint = {2602.01182},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2602.01182},
}
Figure