Adaptive Noise Scheduling for Diffusion-based Imputation of Time-series Single-cell Data

Accurate imputation of missing values in single-cell data remains challenging due to high dimensionality and heterogeneous structures, which can not be effectively handled by most traditional imputation methods. Diffusion-based generative models have shown strong performance by learning data distributions through progressive noising and denoising processes. However, these approaches typically rely on fixed noise schedules, and the impact of different noise scheduling strategies on model performance remains insufficiently understood. Our prior work introduced a time-varying blue noise based diffusion model for imputing missing values in time-series data and demonstrated superior performance over many existing methods. Building on this, we systematically investigate a range of noise scheduling strategies and examine how they influence both the noise injection process and overall imputation performance. Experimental results on single-cell RNA sequencing data show that the model is largely insensitive to the ordering of different noise types during injection, and that different noise schedules have limited impact on reconstruction accuracy, thereby providing greater flexibility in the design of noise schedules during model training. These findings offer a promising direction for advancing generative modeling in high-dimensional data analysis.