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

Researchers introduce FADTI, a diffusion-based framework for multivariate time series imputation that combines Fourier frequency analysis with attention mechanisms to handle missing data in healthcare, traffic, and biological systems. The model demonstrates superior performance over existing methods, particularly when dealing with high missing data rates and distribution shifts.

FADTI addresses a fundamental challenge in applied machine learning: handling missing values in temporal data streams where traditional imputation methods often fail. Multivariate time series imputation remains critical across domains like healthcare monitoring, transportation systems, and scientific research, where sensor malfunctions and sampling irregularities create systematic gaps. Existing Transformer and diffusion models lack explicit mechanisms to capture frequency-domain patterns, leaving them vulnerable when missing data follows structured patterns rather than random distributions.

The framework's innovation centers on the Fourier Bias Projection module, which injects frequency-aware inductive biases into the generative process. This approach recognizes that many real-world signals contain stationary and non-stationary components best understood through spectral analysis. By supporting multiple spectral bases, FADTI can adaptively encode diverse signal characteristics, enabling more robust generalization across different data distributions.

For practitioners and researchers, this advancement has immediate practical implications. High missing rates—common in real-world applications—previously degraded model performance significantly. FADTI's consistent superiority on benchmarks and its newly tested biological dataset suggest improved reliability for critical applications like patient monitoring and ecological tracking. The publicly available code democratizes access to this capability, potentially accelerating adoption across healthcare IT, smart city infrastructure, and scientific computing sectors.

The technical contribution also signals a broader trend: combining classical signal processing insights with modern deep learning architectures yields better inductive biases than purely neural approaches. Stakeholders monitoring AI progress should watch whether frequency-aware designs become standard in temporal modeling, as this could influence development priorities across time series forecasting and anomaly detection applications.

• →FADTI integrates Fourier analysis with diffusion models to improve multivariate time series imputation under high missing data rates
• →The Fourier Bias Projection module enables adaptive encoding of both stationary and non-stationary signal patterns
• →Benchmarks show consistent outperformance compared to state-of-the-art Transformer and diffusion-based methods
• →Open-source code availability facilitates rapid adoption across healthcare, traffic forecasting, and biological modeling applications
• →Combining classical signal processing with modern deep learning demonstrates advantages for temporal data modeling