Blue Noise-based Generative Models for the Imputation of Time-series Data.

Area of Science:

• Machine Learning
• Signal Processing
• Data Science

Background:

• Reconstructing missing high-dimensional time-series data is challenging due to complex dynamics.
• Traditional methods struggle with non-linear relationships and temporal intricacies.
• Generative models, like diffusion methods, show promise but often use white noise, losing frequency details.

Purpose of the Study:

• To develop an improved imputation method for high-dimensional time-series data.
• To address the limitations of white noise in generative diffusion models for time-series imputation.
• To enhance the preservation of frequency-dependent correlations during data reconstruction.

Main Methods:

• Introduced a time-varying blue noise-based conditional score-based diffusion model for imputation (tBN-CSDI).
• Incorporated a time-varying blue noise schedule into the diffusion process.
• Evaluated performance on real-world time-series datasets.

Main Results:

• The tBN-CSDI model demonstrated superior performance compared to conventional white noise-based methods.
• The blue noise schedule effectively preserved fine-scale temporal patterns and frequency-dependent correlations.
• Experimental results confirmed improved imputation accuracy and reliability.

Conclusions:

• The proposed tBN-CSDI model offers a significant advancement in time-series data imputation.
• Time-varying blue noise is a more effective strategy than white noise for diffusion-based imputation.
• Integrating pseudotime analysis with diffusion models presents a promising future research direction for dynamic biological systems.