OpenNovelty.org

Time-Gated Multi-Scale Flow Matching for Time-Series Imputation

ICLR 2026 Conference SubmissionAnonymous Authors
OpenReview Score: 5.6Download Report PDF
Time-series imputationFlow matchingODE-based generative modelsTransformersMulti-scale modeling
Abstract:

We address multivariate time–series imputation by learning the velocity field of a data-conditioned ordinary differential equation (ODE) via flow matching. Our method, Time-Gated Multi-Scale Flow Matching (TG-MSFM), conditions the flow on a structured endpoint comprising observed values, a per-time visibility mask, and short left/right context, processed by a time-aware Transformer whose self-attention is masked to aggregate only from observed timestamps. To recon- cile global trends with local details along the trajectory, we introduce time-gated multi-scale velocity heads on a fixed 1D pyramid and blend them through a time- dependent gate; a mild anti-aliasing filter stabilizes the finest branch. At inference, we use a second-order Heun integrator with a per-step data-consistency projection that keeps observed coordinates exactly on the straight path from the initial noise to the endpoint, reducing boundary artifacts and drift. Training adopts gap-only supervision of the velocity on missing data coordinates, with small optional regu- larizers for numerical stability. Across standard benchmarks, Time-Gated Multi- Scale Flow Matching attains competitive or improved MSE/MAE with favorable speed–quality trade-offs, and ablations isolate the contributions of the time-gated multi-scale heads, masked attention, and the data-consistent ODE integration

Disclaimer
This report is AI-GENERATED using Large Language Models and WisPaper (A scholar search engine). It analyzes academic papers' tasks and contributions against retrieved prior work. While this system identifies POTENTIAL overlaps and novel directions, ITS COVERAGE IS NOT EXHAUSTIVE AND JUDGMENTS ARE APPROXIMATE. These results are intended to assist human reviewers and SHOULD NOT be relied upon as a definitive verdict on novelty.
NOTE that some papers exist in multiple, slightly different versions (e.g., with different titles or URLs). The system may retrieve several versions of the same underlying work. The current automated pipeline does not reliably align or distinguish these cases, so human reviewers will need to disambiguate them manually.
If you have any questions, please contact: mingzhang23@m.fudan.edu.cn

Overview

Overall Novelty Assessment

The paper proposes Time-Gated Multi-Scale Flow Matching (TG-MSFM) for multivariate time-series imputation, combining flow matching with time-aware Transformers, multi-scale velocity heads, and data-consistency projections. It resides in the 'Conditional Flow Matching with Temporal Structure' leaf, which contains only three papers including the original work. This leaf represents a relatively sparse research direction within the broader taxonomy of sixteen papers, suggesting the paper targets a focused niche where flow matching explicitly incorporates temporal dependencies through attention or multi-scale architectures.

The taxonomy reveals neighboring branches exploring alternative generative frameworks (diffusion models, VAEs) and flow-based forecasting methods. The 'Conditional Flow Matching with Temporal Structure' leaf sits under 'Flow Matching Architectures for Time-Series Imputation', which also includes general-purpose flow matching and mask-aware approaches. The sibling papers in this leaf (Flow Matching Dynamics, Conditional Generative Flows) address temporal conditioning but differ in their architectural choices. The taxonomy's scope notes clarify that methods without explicit temporal modeling or those focused solely on forecasting belong elsewhere, positioning this work at the intersection of flow matching and structured temporal reasoning.

Among eleven candidates examined, none clearly refute the three main contributions. The Time-Gated Multi-Scale Flow Matching framework examined ten candidates with zero refutable overlaps, suggesting limited prior work on this specific combination of gating, multi-scale velocity, and flow matching for imputation. The time-gated multi-scale velocity heads contribution examined zero candidates, indicating either a novel architectural element or insufficient search coverage. The Heun integrator with data-consistency projection examined one candidate without refutation, though this limited scope leaves open the possibility of related numerical integration techniques in the broader literature.

Given the sparse taxonomy leaf and limited search scope (eleven candidates from top-K semantic search), the work appears to occupy a relatively unexplored intersection of flow matching, multi-scale architectures, and temporal imputation. However, the analysis does not cover the full landscape of numerical ODE solvers or multi-scale architectures in adjacent domains, so the novelty assessment remains contingent on the examined subset rather than an exhaustive field survey.

Taxonomy

Core-task Taxonomy Papers
16
3
Claimed Contributions
11
Contribution Candidate Papers Compared
0
Refutable Paper

Research Landscape Overview

Core task: multivariate time-series imputation via flow matching. The field addresses the challenge of filling missing values in multivariate temporal data by leveraging continuous normalizing flows and flow matching techniques. The taxonomy reveals four main branches. Flow Matching Architectures for Time-Series Imputation focuses on designing conditional flow models that explicitly incorporate temporal dependencies and missing-data patterns, often through attention mechanisms or recurrent structures. Alternative Generative Frameworks for Time-Series Imputation explores competing paradigms such as variational autoencoders (e.g., Traffic VAE[3]) and diffusion-based methods that offer different trade-offs in sample quality and computational cost. Flow-Based Time-Series Generation and Forecasting extends flow matching beyond imputation to broader generative tasks, including synthesis and prediction (e.g., FlowTS[8], Fm-ts[9]). Theoretical Foundations and Uncertainty Quantification examines the mathematical underpinnings of flow matching, gradient flow dynamics (e.g., Gradient Flow Imputation[4], Wasserstein Gradient Flow[6]), and methods for calibrating predictive uncertainty (e.g., Conformal Flow Prediction[14]). A particularly active line of work centers on conditional flow matching with explicit temporal structure, where methods like Flow Matching Dynamics[2] and Conditional Generative Flows[5] develop architectures that condition on observed time-series segments and propagate information across time steps. TimeGated Flow Matching[0] sits within this cluster, emphasizing gating mechanisms to modulate temporal dependencies during the flow matching process. Compared to Flow Matching Dynamics[2], which explores general dynamical formulations, TimeGated Flow Matching[0] appears to focus more narrowly on learnable gating for selective information flow. Meanwhile, Conditional Generative Flows[5] offers a broader conditional framework that TimeGated Flow Matching[0] may specialize or extend. Open questions in this area include how to balance expressiveness with computational efficiency, how to handle irregular sampling and long-range dependencies, and whether flow matching can robustly quantify imputation uncertainty compared to alternative generative models.

Claimed Contributions

Time-Gated Multi-Scale Flow Matching framework for time-series imputation

The authors introduce TG-MSFM, a method that casts multivariate time-series imputation as learning a data-conditioned ODE via flow matching with visibility-masked self-attention and gap-only supervision. The framework uses a structured endpoint comprising observed values, a per-time visibility mask, and short left/right context, processed by a time-aware Transformer.

10 retrieved papers
Time-gated multi-scale velocity heads with anti-aliasing

The authors propose a velocity decomposition using multi-scale heads on a fixed 1D pyramid that are blended via a time-dependent gate to schedule coarse-to-fine refinement along the ODE trajectory. A light anti-aliasing filter is applied to suppress high-frequency ringing in the finest branch.

0 retrieved papers
Heun integrator with per-step data-consistency projection

The authors pair second-order Heun integration with a per-step data-consistency projection that preserves all observed measurements exactly while evolving unknown entries under the learned dynamics, reducing boundary artifacts and drift during inference.

1 retrieved paper

Core Task Comparisons

Comparisons with papers in the same taxonomy category

Contribution Analysis

Detailed comparisons for each claimed contribution

Contribution

Time-Gated Multi-Scale Flow Matching framework for time-series imputation

The authors introduce TG-MSFM, a method that casts multivariate time-series imputation as learning a data-conditioned ODE via flow matching with visibility-masked self-attention and gap-only supervision. The framework uses a structured endpoint comprising observed values, a per-time visibility mask, and short left/right context, processed by a time-aware Transformer.

Contribution

Time-gated multi-scale velocity heads with anti-aliasing

The authors propose a velocity decomposition using multi-scale heads on a fixed 1D pyramid that are blended via a time-dependent gate to schedule coarse-to-fine refinement along the ODE trajectory. A light anti-aliasing filter is applied to suppress high-frequency ringing in the finest branch.

Contribution

Heun integrator with per-step data-consistency projection

The authors pair second-order Heun integration with a per-step data-consistency projection that preserves all observed measurements exactly while evolving unknown entries under the learned dynamics, reducing boundary artifacts and drift during inference.

Table of Contents

Time-Gated Multi-Scale Flow Matching for Time-Series Imputation | Novelty Validation