Imputation as Inpainting: Diffusion models for SpatioTemporal Data Imputation

Taeyoung Yun, Haewon Jung, Jiwoo Son

Keywords: Spatiotemporal Data Imputatation, Diffusion Models, Inpainting, Traffic forecasting

TL;DR: Propose a novel framework to impute missing values in spatiotemporal data using graph-structured diffusion models

Abstract: Spatiotemporal data mining plays a crucial role in real-world scenarios such as air quality monitoring and intelligent traffic management. However, real-world spatiotemporal data collected in such scenarios is often incomplete due to sensor failures or transmission loss. Spatiotemporal imputation aims to fill in the missing values based on the observed values and their underlying spatiotemporal dependence. Previous dominant imputation models relied on autoregressive methods, which suffered from error accumulation. To overcome this limitation, emerging generative models like diffusion probabilistic models (DPM) can be employed for imputing missing values. These models are conditioned on observations to avoid relying solely on inaccurate historical imputation methods. However, applying diffusion models to spatiotemporal imputation presents challenges, particularly in extracting and utilizing conditional information from observed data. In this paper, we propose a novel framework for utilizing diffusion models for spatiotemporal imputation, by formulating imputation as inpainting problem. We first train an unconditional diffusion model for predicting the whole spatiotemporal data. To condition the generative process, we follow the scheme proposed by RePaint \citep{lugmayr2022repaint}, only alter reverse diffusion iterations by sampling the unobserved regions using the observed data information. To model spatial dependencies, we utilize a GNN-based backbone for DPM. We compare our model with state-of-the-art baselines in various missing patterns of two real-world spatiotemporal benchmark datasets.