Inductive Spatial Temporal Prediction Under Data Drift with Informative Graph Neural Network

TL;DR

This work tackles inductive spatial temporal prediction under data drift, where external events and expanding entity sets shift data distributions. It introduces INF-GNN, which distills diversified invariant patterns through a novel Relation Importance ($RI$) metric to form an informative subgraph and uses an informative temporal memory buffer to emphasize influential timestamps; both are integrated via RI loss optimization that combines standard loss with Elastic Weight Consolidation and RI-based regularization. On the PEMS3-Stream traffic dataset, INF-GNN achieves state-of-the-art performance for both existing and newly added nodes under substantial drift, demonstrating improved generalization and robustness in dynamic graphs. The approach offers interpretable pattern consolidation and practical applicability to rapidly changing spatio-temporal systems.

Abstract

Inductive spatial temporal prediction can generalize historical data to predict unseen data, crucial for highly dynamic scenarios (e.g., traffic systems, stock markets). However, external events (e.g., urban structural growth, market crash) and emerging new entities (e.g., locations, stocks) can undermine prediction accuracy by inducing data drift over time. Most existing studies extract invariant patterns to counter data drift but ignore pattern diversity, exhibiting poor generalization to unseen entities. To address this issue, we design an Informative Graph Neural Network (INF-GNN) to distill diversified invariant patterns and improve prediction accuracy under data drift. Firstly, we build an informative subgraph with a uniquely designed metric, Relation Importance (RI), that can effectively select stable entities and distinct spatial relationships. This subgraph further generalizes new entities' data via neighbors merging. Secondly, we propose an informative temporal memory buffer to help the model emphasize valuable timestamps extracted using influence functions within time intervals. This memory buffer allows INF-GNN to discern influential temporal patterns. Finally, RI loss optimization is designed for pattern consolidation. Extensive experiments on real-world dataset under substantial data drift demonstrate that INF-GNN significantly outperforms existing alternatives.

Figures (7)

• Figure 1: Motivating experiments. (a) Entities with invariant patterns will have distribution remain stable over time. (b) Certain timestamps will have considerable deviation gaps across different time intervals.
• Figure 2: General framework of INF-GNN. (a) Using RI metric to select nodes that are stable and have little mutual information with their neighbors to construct informative subgraph, which are further used for simulation of new entities. (b) Selecting informative timestamps by influence function to jointly train with all timestamps. (c) A simple surrogate spatial temporal predicting model is adapted with RI loss optimization to make predictions.
• Figure 3: Prediction accuracy comparison across different length of time interval
• Figure 4: Impact of three main components.
• Figure 5: Visualization of components. (a) Informative subgraph contains entities with stable, but high deviation distribution reflecting its invariant and informative characteristic. (b) Red vertical value indicates the frequency of stamps being selected to informative temporal memory buffer. Those with bigger deviation gap will be more frequently selected

Theorems & Definitions (4)

• definition thmcounterdefinition
• definition thmcounterdefinition
• definition thmcounterdefinition
• definition thmcounterdefinition