Enhancing Dynamic GCN for Node Attribute Forecasting with Meta Spatial-Temporal Learning (Student Abstract)

Bo Wu, Xun Liang, Xiangping Zheng, Jun WangAAAI 2023 · DOI

Abstract

Predicting the evolution of node attributes in dynamic networks is a critical task with applications ranging from traffic management to epidemiology. Traditional Dynamic Graph Convolutional Networks (DGCNs) often struggle to capture the irregular and complex spatial-temporal interactions inherent in real-world data. This paper introduces meta-DGCN, a novel framework that treats spatial relationships and temporal changes as meta-learning tasks. By incorporating a gated encoding layer and a dual-task meta-optimization strategy, meta-DGCN demonstrates superior adaptability and forecasting accuracy. Experimental results on three real-world datasets—WikiMaths, EnglandCovid, and PedalMe—show that meta-DGCN consistently outperforms state-of-the-art baselines, particularly in highly volatile scenarios.

Introduction

Node attribute forecasting involves predicting the future states of nodes within a graph structure that evolves over time. Examples include forecasting traffic flow on road networks or tracking infection rates across geographical regions. While existing models combine Graph Convolutional Networks (GCNs) with sequence models like Long Short-Term Memory (LSTM) networks, they often assume relatively stable or predictable patterns of change.

However, real-world dynamic graphs are characterized by "messy" and irregular interactions. The relationship between a node and its neighbors (spatial) and the evolution of a node's state over time (temporal) are deeply intertwined and frequently shifting. Current methodologies lack the flexibility to adapt to these rapid changes, leading to suboptimal forecasting performance.

Methodology: The meta-DGCN Framework

The proposed meta-DGCN framework addresses these challenges through a three-tiered architecture designed for rapid adaptation.

1. Encoding Layer with Gating Mechanism

The foundation of the model is an encoding layer that utilizes a standard GCN to generate node embeddings. To better manage the flow of information across time steps, the authors integrate a sophisticated gating mechanism inspired by LSTMs. This allows the model to selectively retain or discard information from previous states and the current graph topology, ensuring that the node representations remain relevant to the most recent dynamics.

2. Spatial-Temporal Task Construction

A key innovation of meta-DGCN is its approach to task construction. Rather than training on the dataset as a monolithic sequence, the framework decomposes the learning process into specific meta-tasks:

  • Spatial Tasks: These tasks involve sampling edges to understand the underlying connectivity and influence between nodes.
  • Temporal Tasks: These tasks involve sampling nodes to capture how their attributes fluctuate over time. Each task is structured with a "support set" for local adaptation and a "query set" for evaluating the model's generalization capability.

3. Meta-Optimization

The model employs meta-learning to optimize its parameters. Instead of merely learning to fit the training data, meta-DGCN learns how to learn from new patterns. By minimizing a combined adaptation loss that accounts for both spatial and temporal performance, the model becomes highly sensitive to the evolving trends within the dynamic graph. This meta-optimization allows the model to quickly adjust its weights when faced with irregular data shifts.

Experimental Evaluation

The performance of meta-DGCN was evaluated against several state-of-the-art baselines, including GConvGRU, DCRNN, and EvolveGCN, across three diverse real-world datasets:

  • WikiMaths: Traffic patterns on Wikipedia pages related to mathematics.
  • EnglandCovid: Daily COVID-19 infection rates across various regions in England.
  • PedalMe: Demand and flow data from a bicycle delivery service.

Results and Discussion

The experimental results, summarized in the table below, indicate that meta-DGCN achieves the lowest Mean Squared Error (MSE) across all benchmarks.

Comparison of MSE results for meta-DGCN and baseline models across three datasets.

The most significant improvement was observed in the EnglandCovid dataset. Meta-DGCN achieved an MSE of 0.5411, a substantial reduction compared to the next best baseline, DCRNN, which recorded 0.8323. This suggests that the meta-learning approach is exceptionally well-suited for highly volatile and non-linear data, such as the spread of a pandemic. In the WikiMaths and PedalMe datasets, meta-DGCN also demonstrated superior performance, confirming its robustness across different types of dynamic networks.

Conclusion

Meta-DGCN represents a significant step forward in node attribute forecasting by leveraging meta-learning to handle the complexities of dynamic graphs. By treating spatial and temporal dynamics as adaptive tasks, the model achieves a level of flexibility that traditional DGCNs lack. While this research demonstrates the potential of meta-learning in this domain, future work will focus on scaling the framework for larger graphs and exploring more complex task construction strategies to further enhance adaptation speed and accuracy.

Summary updated

BibTeX
@inproceedings{wu2023enhancing,
  author = {Bo Wu and Xun Liang and Xiangping Zheng and Jun Wang},
  title = {{Enhancing Dynamic GCN for Node Attribute Forecasting with Meta Spatial-Temporal Learning (Student Abstract)}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2023},
  doi = {10.1609/aaai.v37i13.27040},
}
Figure