GradDiff-N-Tab: Gradient Noise Tabular Data Diffusion Model Imputation

Ari Wibisono, Petrus Mursanto, Denny, Simon See

Keywords: Tabular data, Imputation, Cosine Scheduler, Stable Diffusion Models, Gradient Noise

TL;DR: In this research, the author propose an imputation method in the diffusion models algorithm for non-normal or extreme data. The improvement process involves noise generation, loss function, and noise scheduler changes.

Abstract: Imputation is one of the methods to improve the quality of a dataset. The imputation problem can be solved using statistical techniques, machine learning algorithms, and generative models. This research proposes improving the standard imputation algorithm based on the Diffusion Model. We propose to use Perlin noise generation to generate noise at each step of the diffusion mode and propose a scheduler to improve the performance of the diffusion model-based imputation algorithm. Perlin noise generation and cosine scheduler have a positive influence on improving the performance of non-normal data imputation. Four real-world datasets are used to evaluate our proposed methods. Based on the evaluation tests of the RMSE value, our proposed method produces a 10% lower RMSE value than the baseline imputation algorithms based on diffusion models, GAN, and VAE.

Submission Number: 3