Hi-Patch: Hierarchical Patch GNN for Irregular Multivariate Time Series Modeling

Yicheng Luo, Bowen Zhang, Zhen Liu, Qianli Ma

Keywords: irregular multivariate time series, graph neural network

Abstract:

Multi-scale information is crucial for multivariate time series modeling. However, most existing time series multi-scale analysis methods treat all variables in the same manner, which is not well adaptive to Irregular Multivariate Time Series (IMTS), where different variables have distinct original scales/sampling rates. Therefore, extracting temporal and inter-variable dependencies at multiple scales
in IMTS remains challenging. To fill this gap, we propose a hierarchical patch graph network Hi-Patch. The key components of Hi-Patch are an intra-patch graph layer and several inter-patch graph layers. The intra-patch graph layer flexibly represents and fully captures both the local temporal and inter-variable dependencies of densely sampled variables at the original scale by employing fully connected graph networks within each patch, and obtains patch-level node representations through aggregation. Subsequently, several inter-patch graph layers are stacked to form a hierarchical architecture, where each layer updates specific patch-level nodes through scale-specific graph networks, progressively representing and extracting more global temporal and inter-variable features of both sparsely and densely sampled variables, and further aggregating to produce the next patch-level node representations. The output of the last inter-patch graph
layer is fed into task-specific decoders to adapt to different downstream tasks. Experimental results on 8 datasets show that Hi-Patch outperforms a range of state-of-the-art models in both IMTS forecasting and classification tasks. Code is available at this repository: https://anonymous.4open.science/r/Hi-Patch-F42E.

Primary Area: applications to computer vision, audio, language, and other modalities

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: 2938