G-TIGRE: A new generative framework for Multivariate Time Series Imputation By Graph Neural Networks

TL;DR — We introduce a novel framework that combines Generative Adversarial Networks (GANs) and Graph Neural Networks (GNNs) to address the challenge of missing data in multivariate time series

Abstract

The persistent challenge of handling missing values in multivariate time series (MTS) data demands precise solutions to avoid potential pitfalls in real-world applications. Conventional imputation methods often struggle to capture effective spatio-temporal representations of such data, failing to exploit its intrinsic temporal nature and intricate inter-variable relationships. In recent years, deep learning-based imputation methods have gained popularity. However, they often lack dedicated structures and models specifically designed to address this unique challenge. In response to these challenges, we introduce a novel framework called G-TIGRE, which synergistically leverages the capabilities of two prominent research streams in this field: Generative Adversarial Networks (GANs) and Graph Neural Networks (GNNs). GANs excel at effectively modeling data distributions, while GNNs demonstrate remarkable proficiency in extracting spatio-temporal features from data. By integrating these two techniques, which have not previously been explored together in this domain, G-TIGRE addresses several critical issues, including the elimination of the need to make assumptions about data stationarity, the ability to train with incomplete data, and the enhancement of spatio-temporal representation learning. Through extensive experiments conducted on a diverse benchmark of state-of-the-art methods, we establish that G-TIGRE achieves competitive performance, closely rivaling the top-performing models. Furthermore, an in-depth ablation study sheds light on the unique contributions of each component within G-TIGRE, elucidating its effectiveness in MTS imputation. This work introduces an exciting shift in addressing the persistent challenges of missing data in multivariate time series, with far-reaching implications across various domains.