Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

STRGCN: Capturing Asynchronous Spatio-Temporal Dependencies for Irregular Multivariate Time Series Forecasting

Yulong Wang, Xiaofeng Hu, Xiaojian Cui, Kai Wang

TL;DR

This work tackles irregular multivariate time series forecasting by avoiding costly pre-alignment and instead modeling all observations as nodes in a fully connected graph. The STRGCN framework uses a Spatio-Temporal Relational Graph Convolution that decouples temporal and spatial edge components, enabling robust handling of asynchronous dependencies, while a hierarchical Sandwich structure aggregates information to balance local detail with global context and reduce computation. Empirical results across four public IMTS datasets show STRGCN achieving state-of-the-art MSE and competitive MAE, with notable memory and training-speed advantages driven by the Sandwich and low-rank edge modeling. The approach advances practical forecasting in domains with heterogeneous sensor frequencies and irregular observations, offering scalable, expressive modeling of complex spatio-temporal relationships.

Abstract

Irregular multivariate time series (IMTS) are prevalent in real-world applications across many fields, where varying sensor frequencies and asynchronous measurements pose significant modeling challenges. Existing solutions often rely on a pre-alignment strategy to normalize data, which can distort intrinsic patterns and escalate computational and memory demands. Addressing these limitations, we introduce STRGCN, a Spatio-Temporal Relational Graph Convolutional Network that avoids pre-alignment and directly captures the complex interdependencies in IMTS by representing them as a fully connected graph. Each observation is represented as a node, allowing the model to effectively handle misaligned timestamps by mapping all inter-node relationships, thus faithfully preserving the asynchronous nature of the data. Moreover, we enhance this model with a hierarchical ``Sandwich'' structure that strategically aggregates nodes to optimize graph embeddings, reducing computational overhead while maintaining detailed local and global context. Extensive experiments on four public datasets demonstrate that STRGCN achieves state-of-the-art accuracy, competitive memory usage and training speed.

STRGCN: Capturing Asynchronous Spatio-Temporal Dependencies for Irregular Multivariate Time Series Forecasting

TL;DR

This work tackles irregular multivariate time series forecasting by avoiding costly pre-alignment and instead modeling all observations as nodes in a fully connected graph. The STRGCN framework uses a Spatio-Temporal Relational Graph Convolution that decouples temporal and spatial edge components, enabling robust handling of asynchronous dependencies, while a hierarchical Sandwich structure aggregates information to balance local detail with global context and reduce computation. Empirical results across four public IMTS datasets show STRGCN achieving state-of-the-art MSE and competitive MAE, with notable memory and training-speed advantages driven by the Sandwich and low-rank edge modeling. The approach advances practical forecasting in domains with heterogeneous sensor frequencies and irregular observations, offering scalable, expressive modeling of complex spatio-temporal relationships.

Abstract

Irregular multivariate time series (IMTS) are prevalent in real-world applications across many fields, where varying sensor frequencies and asynchronous measurements pose significant modeling challenges. Existing solutions often rely on a pre-alignment strategy to normalize data, which can distort intrinsic patterns and escalate computational and memory demands. Addressing these limitations, we introduce STRGCN, a Spatio-Temporal Relational Graph Convolutional Network that avoids pre-alignment and directly captures the complex interdependencies in IMTS by representing them as a fully connected graph. Each observation is represented as a node, allowing the model to effectively handle misaligned timestamps by mapping all inter-node relationships, thus faithfully preserving the asynchronous nature of the data. Moreover, we enhance this model with a hierarchical ``Sandwich'' structure that strategically aggregates nodes to optimize graph embeddings, reducing computational overhead while maintaining detailed local and global context. Extensive experiments on four public datasets demonstrate that STRGCN achieves state-of-the-art accuracy, competitive memory usage and training speed.
Paper Structure (21 sections, 21 equations, 4 figures, 2 tables)

This paper contains 21 sections, 21 equations, 4 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Irregular Multivariate Time Series Forecasting
  4. Graph Based Spatio-Temporal Modeling
  5. Problem Definition
  6. Method
  7. Overall Architecture
  8. Fully Connected Graph Representation of IMTS
  9. Spatio-Temporal Relational Graph Convolution
  10. Temporal and Spatial Relation Decomposition
  11. Normalization Factor
  12. Final STRGCN Update Rule
  13. Hierarchical Sandwich Structure
  14. Experiments
  15. Experimental Details
  16. ...and 6 more sections

Figures (4)

  • Figure 1: (a) Illustrates the characteristics of an irregular multivariate time series, where variable $v_1$ has irregular sampling intervals, and $v_1$, $v_2$, and $v_3$ have different sampling frequencies and misaligned timestamps. (b) displays the pre-alignment representation for IMTS, which results in significant additional memory consumption due to the inclusion of padded values. (c) shows the fully-connected graph representation for IMTS, where each observation is treated as a node, and some edges are omitted for clarity.
  • Figure 2: Overview of STRGCN, consisting of the following key components: (a) the Fully-Connected Graph Transformation Module, which converts IMTS data into a compact representation of a fully connected graph; (b) the Spatio-Temporal Relational Graph Convolution Layer, designed to capture asynchronous spatio-temporal dependencies; and (c) the Hierarchical Sandwich Structure, which integrates local and global semantic relationships while mitigating computational complexity.
  • Figure 3: Illustration of the Hyper-nodes Generation and the Hierarchical Sandwich Structure. Panel (a) shows the generation of hyper-nodes by uniformly sampling nodes along the temporal axis for each variable with a predefined window length. Panel (b) provides an intuitive understanding of the bottom, middle, and top layers in the Sandwich structure.
  • Figure 4: Performance Analysis of STRGCN: Assessing Average MSE, Training Time, and Memory Usage (lower is better), evaluated on the PhysioNet Dataset.