tBN-CSDI: a time-varying blue noise-based diffusion model for time-series imputation.

Area of Science:

• Biomedical data analysis
• Computational biology
• Machine learning for time-series data

Background:

• Missing data imputation is challenging in high-dimensional time-series analysis.
• Traditional methods fail to capture complex nonlinear dependencies.
• Existing diffusion models use isotropic white noise, obscuring frequency-dependent correlations.

Purpose of the Study:

• To introduce a novel imputation method, time-varying blue noise-based conditional score-based diffusion model (tBN-CSDI).
• To improve the recovery of high-frequency temporal patterns in sparse time-series data.
• To enhance the accuracy of missing data imputation in biomedical and biological datasets.

Main Methods:

• Developed a time-varying blue noise-based conditional score-based diffusion model (tBN-CSDI).
• Modulated the noise schedule based on data frequency characteristics.
• Applied the model to healthcare and single-cell RNA-seq datasets.

Main Results:

• tBN-CSDI consistently outperforms existing imputation methods.
• Achieved over a 30% reduction in imputation error under high data sparsity.
• Demonstrated improved recovery of subtle, high-frequency temporal patterns.

Conclusions:

• tBN-CSDI is a robust and effective solution for imputing sparse and noisy time-series data.
• The method shows potential for improving change-point detection and gene regulatory network inference.
• Code and data are publicly available on GitHub for broader research application.