Injecting Global Context for Multivariate Time Series Forecasting on Variable Subsets

Abstract

Variable Subset Forecasting (VSF) presents a critical challenge where variable availability fluctuates between the training and inference stages. This paper introduces a novel solution to the underexplored VSF problem, seamlessly integrating with existing variate-wise attention architectures. Our approach features two key innovations: (1) a global context-enriched attention mechanism maintaining a holistic data perspective with incomplete variable sets, and (2) a random group mechanism simulating missing data scenarios during training. These innovations effectively address the challenges of maintaining variable correlations under variable stochastic absence and adapting to dynamic input variate without retraining. Experiments demonstrate that our approach improves predictive performance by an average of 16%, with gains of up to 40% in particularly challenging scenarios involving significant variable absence, while adding only a 1% computational overhead. This substantial performance boost, coupled with adaptability to various architectures, positions our approach as a robust solution for real-world MTSF applications facing variable unavailability and domain shifts.