FADTI: Fourier and Attention Driven Diffusion for Multivariate Time Series Imputation

2025-12-17

AI Summary

Multivariate time series imputation faces structural missingness challenges arising from sensor failures and irregular sampling. Existing Transformer- and diffusion-based methods lack frequency-awareness and inductive biases, limiting generalization. This paper proposes F-Diffusion—the first diffusion framework integrating a learnable Fourier Bias Projection (FBP) into the denoising process, explicitly injecting frequency-domain priors to adaptively model both stationary and non-stationary dynamics. It synergistically combines temporal self-attention, gated convolutions, and multi-basis spectral representations to enhance robustness against diverse missing patterns and distributional shifts. Extensive experiments on multiple established benchmarks and a newly curated biological time series dataset demonstrate that F-Diffusion significantly outperforms state-of-the-art methods, maintaining high imputation accuracy even under extreme missing rates exceeding 70%.

Technology Category

• Machine Learning: Time-Series/Data Streams
• Computer Vision: Diffusion Models for Vision
• Intelligent Robots: Multimodal Perception & Sensor Fusion

Application Category

• Graph Algorithms and Modeling for the Web: Representation, reconstruction, and subgraph or motif discovery in Web-related graphs
• User Modeling, Personalization and Recommendation: Federated recommendation systems and personalization
• Web Mining and Content Analysis: Content-based information diffusion

Abstract

Multivariate time series imputation is fundamental in applications such as healthcare, traffic forecasting, and biological modeling, where sensor failures and irregular sampling lead to pervasive missing values. However, existing Transformer- and diffusion-based models lack explicit inductive biases and frequency awareness, limiting their generalization under structured missing patterns and distribution shifts. We propose FADTI, a diffusion-based framework that injects frequency-informed feature modulation via a learnable Fourier Bias Projection (FBP) module and combines it with temporal modeling through self-attention and gated convolution. FBP supports multiple spectral bases, enabling adaptive encoding of both stationary and non-stationary patterns. This design injects frequency-domain inductive bias into the generative imputation process. Experiments on multiple benchmarks, including a newly introduced biological time series dataset, show that FADTI consistently outperforms state-of-the-art methods, particularly under high missing rates. Code is available at https://anonymous.4open.science/r/TimeSeriesImputation-52BF

Problem

• Addresses missing values in multivariate time series from sensor failures and irregular sampling.
• Improves generalization under structured missing patterns and distribution shifts.
• Enhances frequency awareness and inductive biases in diffusion-based imputation models.

Innovation

• Fourier Bias Projection module for frequency-aware feature modulation
• Combines self-attention and gated convolution for temporal modeling
• Adaptive encoding of stationary and non-stationary patterns via spectral bases