Denoising Diffusion Probabilistic Imputation Model for Tabular Data

Data integrity is often compromised when information is stored, resulting in missing values within tabular datasets. The task of accurately recovering these lost values has emerged as a vibrant area of research. This study introduces a novel generative model termed the Denoising Diffusion Probabilistic Imputation Model (DDPIM) that is designed to predict missing data values with both efficiency and precision. DDPIM oper-ates as a diffusion probability model, discerning the true data distribution by first introducing Gaussian noise into clean data and then denoising it during the reverse process. The model commences by training a denoising diffusion model on a complete dataset, thereby acquiring the distribution characteristics of data. Subsequently, this pre-trained model is employed to reconstruct data points that are incomplete, i.e., those that contain missing values. DDPIM utilizes the distribution learned from the diffusion model as prior information and incorporates the known attributes within the incomplete samples as guiding conditions. This approach guides the data points with missing values towards the true data distribution. By leveraging the influence of the known attributes during the denoising phase, the incomplete samples are ultimately aligned with the actual data distribution. Extensive experiments conducted on a wide array of medical datasets demonstrate that our method yields superior performance compared to existing approaches.