ASTGI: Adaptive Spatio-Temporal Graph Interactions for Irregular Multivariate Time Series Forecasting

TL;DR

This paper tackles irregular multivariate time series forecasting by introducing ASTGI, an adaptive spatio-temporal graph framework. It preserves raw information through direct spatio-temporal point representations and replaces fixed interaction structures with per-point neighborhood graphs learned in a dedicated embedding space. A relation-aware dynamic propagation mechanism updates features across adaptive graphs, and a query-point fusion head enables accurate predictions for new timestamps and variables. Empirical results on multiple public benchmarks demonstrate state-of-the-art performance and robust generalization, underscoring the practicality of adaptive graph interactions for irregular time series data.

Abstract

Irregular multivariate time series (IMTS) are prevalent in critical domains like healthcare and finance, where accurate forecasting is vital for proactive decision-making. However, the asynchronous sampling and irregular intervals inherent to IMTS pose two core challenges for existing methods: (1) how to accurately represent the raw information of irregular time series without introducing data distortion, and (2) how to effectively capture the complex dynamic dependencies between observation points. To address these challenges, we propose the Adaptive Spatio-Temporal Graph Interaction (ASTGI) framework. Specifically, the framework first employs a Spatio-Temporal Point Representation module to encode each discrete observation as a point within a learnable spatio-temporal embedding space. Second, a Neighborhood-Adaptive Graph Construction module adaptively builds a causal graph for each point in the embedding space via nearest neighbor search. Subsequently, a Spatio-Temporal Dynamic Propagation module iteratively updates information on these adaptive causal graphs by generating messages and computing interaction weights based on the relative spatio-temporal positions between points. Finally, a Query Point-based Prediction module generates the final forecast by aggregating neighborhood information for a new query point and performing regression. Extensive experiments on multiple benchmark datasets demonstrate that ASTGI outperforms various state-of-the-art methods.

Figures (6)

• Figure 1: An illustration of information distortion in mainstream IMTSF paradigms. (a) Raw IMTS. (b) Interpolation-based: Converts irregular series into equally spaced series through numerical interpolation. (c) Time-aligned-based: Maps the observations of all variables to a unified timeline and fills in missing values. (d) Patch-aligned-based: Slices the time series into multiple patches.
• Figure 2: From fixed interaction rules to adaptive graph interaction. (a) Raw IMTS. (b) ODE-based interaction: Follows the temporal sequence to model continuous dynamics between observations. (c) Static Structure Interaction: Employs a fixed, predefined graph structure, confining information exchange to a static set of connections. (d) ASTGI (Ours): Adaptively constructs a unique graph for each observation point, enabling the capture of complex and dynamic dependencies.
• Figure 3: Overview of the ASTGI framework. (a) Directly representing each discrete observation as a spatio-temporal point. (b) Adaptively constructing a causal graph for each point. (c) Iteratively propagating information on the adaptive graphs to update features. (d) Unifying prediction as a neighborhood aggregation task for a query point.
• Figure 4: Parameter sensitivity studies of main hyper-parameters in ASTGI.
• Figure 5: Visualization of the adaptively learned causal graph. The plot displays the interactions between observation points for a sample from the MIMIC dataset. The x-axis represents time (hours), and the y-axis represents different variables. Arrows indicate the direction of information flow (from history to query). The model successfully captures (1) synchronous correlations between variables (e.g., Bilirubin Direct and pH at $t=8$), (2) long-range temporal dependencies (e.g., Troponin I self-connection $t=4 \rightarrow t=7$), and (3) cross-variable lagged effects. This confirms that ASTGI adaptively constructs a sparse and meaningful interaction topology.