Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

STRAP: Spatio-Temporal Pattern Retrieval for Out-of-Distribution Generalization

Haoyu Zhang, Wentao Zhang, Hao Miao, Xinke Jiang, Yuchen Fang, Yifan Zhang

TL;DR

Extensive experiments show that STRAP consistently outperforms state-of-the-art STGNN baselines on STOOD tasks, demonstrating its robustness, adaptability, and strong generalization capability without task-specific fine-tuning.

Abstract

Spatio-Temporal Graph Neural Networks (STGNNs) have emerged as a powerful tool for modeling dynamic graph-structured data across diverse domains. However, they often fail to generalize in Spatio-Temporal Out-of-Distribution (STOOD) scenarios, where both temporal dynamics and spatial structures evolve beyond the training distribution. To address this problem, we propose an innovative Spatio-Temporal Retrieval-Augmented Pattern Learning framework,STRAP, which enhances model generalization by integrating retrieval-augmented learning into the STGNN continue learning pipeline. The core of STRAP is a compact and expressive pattern library that stores representative spatio-temporal patterns enriched with historical, structural, and semantic information, which is obtained and optimized during the training phase. During inference, STRAP retrieves relevant patterns from this library based on similarity to the current input and injects them into the model via a plug-and-play prompting mechanism. This not only strengthens spatio-temporal representations but also mitigates catastrophic forgetting. Moreover, STRAP introduces a knowledge-balancing objective to harmonize new information with retrieved knowledge. Extensive experiments across multiple real-world streaming graph datasets show that STRAP consistently outperforms state-of-the-art STGNN baselines on STOOD tasks, demonstrating its robustness, adaptability, and strong generalization capability without task-specific fine-tuning.

STRAP: Spatio-Temporal Pattern Retrieval for Out-of-Distribution Generalization

TL;DR

Extensive experiments show that STRAP consistently outperforms state-of-the-art STGNN baselines on STOOD tasks, demonstrating its robustness, adaptability, and strong generalization capability without task-specific fine-tuning.

Abstract

Spatio-Temporal Graph Neural Networks (STGNNs) have emerged as a powerful tool for modeling dynamic graph-structured data across diverse domains. However, they often fail to generalize in Spatio-Temporal Out-of-Distribution (STOOD) scenarios, where both temporal dynamics and spatial structures evolve beyond the training distribution. To address this problem, we propose an innovative Spatio-Temporal Retrieval-Augmented Pattern Learning framework,STRAP, which enhances model generalization by integrating retrieval-augmented learning into the STGNN continue learning pipeline. The core of STRAP is a compact and expressive pattern library that stores representative spatio-temporal patterns enriched with historical, structural, and semantic information, which is obtained and optimized during the training phase. During inference, STRAP retrieves relevant patterns from this library based on similarity to the current input and injects them into the model via a plug-and-play prompting mechanism. This not only strengthens spatio-temporal representations but also mitigates catastrophic forgetting. Moreover, STRAP introduces a knowledge-balancing objective to harmonize new information with retrieved knowledge. Extensive experiments across multiple real-world streaming graph datasets show that STRAP consistently outperforms state-of-the-art STGNN baselines on STOOD tasks, demonstrating its robustness, adaptability, and strong generalization capability without task-specific fine-tuning.

Paper Structure

This paper contains 44 sections, 1 theorem, 31 equations, 10 figures, 8 tables, 2 algorithms.

Table of Contents

  1. Introduction
  2. Related Work
  3. Continue Learning on STOOD Tasks
  4. Pattern Retrieval Learning
  5. Preliminaries
  6. Definition 1. (Dynamic Spatio-Temporal Graph)
  7. Definition 2. (Spatio-Temporal Out-of-Distribution Learning)
  8. Definition 3. (Unified STOOD Prediction Framework)
  9. StRap Framework
  10. Spatial Temporal Subgraph Chunking
  11. Multi-dimensional Pattern Library Construction
  12. Pattern Key Generation
  13. Pattern Value Generation
  14. Pattern Retrieval and Knowledge Fusion
  15. Pattern Library Retrieval
  16. ...and 29 more sections

Key Result

Theorem B.1

Let $X$ be the original data. Consider two feature extraction approaches: The mutual information advantage of the decomposed approach satisfies:

Figures (10)

  • Figure 1: The overall framework of StRap.
  • Figure 2: Impact of different pattern libraries and keys. Left: library. Right: key.
  • Figure 3: Test set distributions for ENERGY-Wind for 4 periods 0 (a), 1 (b), 2 (c), and 3 (d).
  • Figure 4: Performance analysis (MAPE) of different horizons across various baselines.
  • Figure 5: Performance analysis of different backbone across various fusion ratios $\gamma$.
  • ...and 5 more figures

Theorems & Definitions (2)

  • Theorem B.1
  • proof