Search papers, labs, and topics across Lattice.
This paper investigates the role of Reversible Instance Normalization (RevIN) in time series forecasting, highlighting challenges related to temporal, spatial, and conditional distribution shifts. Through ablation studies, the authors identify redundant and detrimental components within the RevIN architecture. Based on these findings, they propose new perspectives to enhance RevIN's robustness and generalization capabilities.
RevIN, a popular normalization technique for time series forecasting, can be significantly improved by removing redundant components, challenging current best practices.
Data normalization is a crucial component of deep learning models, yet its role in time series forecasting remains insufficiently understood. In this paper, we identify three central challenges for normalization in time series forecasting: temporal input distribution shift, spatial input distribution shift, and conditional output distribution shift. In this context, we revisit the widely used Reversible Instance Normalization (RevIN), by showing through ablation studies that several of its components are redundant or even detrimental. Based on these observations, we draw new perspectives to improve RevIN's robustness and generalization.
