GCGNet: Graph-Consistent Generative Network for Time Series Forecasting with Exogenous Variables
A novel approach known as Graph-Consistent Generative Network (GCGNet) has been introduced for forecasting time series that include exogenous variables. These exogenous variables enhance the prediction of future endogenous variables, necessitating the modeling of both temporal correlations (dependencies from past to future) and channel correlations (the effect of exogenous variables on endogenous ones). Traditional methods generally adopt a two-step approach, treating these correlations separately, which restricts their capacity to capture joint correlations across time and channels. GCGNet overcomes this challenge by simultaneously modeling both types of correlations within a graph-consistent generative framework. Furthermore, since real-world time series often experience noise, ensuring robustness is essential. The paper can be found on arXiv under ID 2603.08032.
Key facts
- GCGNet is a graph-consistent generative network for time series forecasting with exogenous variables.
- Exogenous variables offer valuable supplementary information for predicting future endogenous variables.
- Forecasting must consider both temporal correlations and channel correlations.
- Existing methods use a two-step strategy, modeling temporal and channel correlations separately.
- GCGNet captures joint correlations across time and channels.
- Real-world time series are frequently affected by various forms of noise.
- Robustness is critical in correlation modeling.
- The paper is available on arXiv:2603.08032.
Entities
Institutions
- arXiv