GraphSTAGE: Channel-Preserving Graph Neural Networks for Time Series Forecasting
Keywords: Time Series Forecasting, Graph Neural Networks, Channel-Preserving
TL;DR: GraphSTAGE is a purely graph neural network based model that decouples intra-series and inter-series dependencies learning, using a channel-preserving framework to reduce noise and improve performance in multivariate time series forecasting.
Abstract: Recent advancements in multivariate time series forecasting (MTSF) have increasingly focused on the core challenge of learning dependencies within sequences, specifically intra-series (temporal), inter-series (spatial), and cross-series dependencies. While extracting multiple types of dependencies can theoretically enhance the richness of learned correlations, it also increases computational complexity and may introduce additional noise. The trade-off between the variety of dependencies extracted and the potential interference has not yet been fully explored. To address this challenge, we propose GraphSTAGE, a purely graph neural network (GNN)-based model that decouples the learning of intra-series and inter-series dependencies. GraphSTAGE features a minimal architecture with a specially designed embedding and patching layer, along with the STAGE (Spatial-Temporal Aggregation Graph Encoder) blocks. Unlike channel-mixing approaches, GraphSTAGE is a channel-preserving method that maintains the shape of the input data throughout training, thereby avoiding the interference and noise typically caused by channel blending. Extensive experiments conducted on 13 real-world datasets demonstrate that our model achieves performance comparable to or surpassing state-of-the-art methods. Moreover, comparative experiments between our channel-preserving framework and channel-mixing designs show that excessive dependency extraction and channel blending can introduce noise and interference. As a purely GNN-based model, GraphSTAGE generates learnable graphs in both temporal and spatial dimensions, enabling the visualization of data periodicity and node correlations to enhance model interpretability.
Primary Area: learning on time series and dynamical systems
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 4428
Loading