Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

TempoKGAT: A Novel Graph Attention Network Approach for Temporal Graph Analysis

Lena Sasal, Daniel Busby, Abdenour Hadid

TL;DR

TempoKGAT addresses forecasting on temporal graphs by integrating time-decaying weights with selective top-k neighbor aggregation into a Graph Attention Network. The model applies temporal decay to node features and restricts neighbor aggregation to the most informative neighbors based on edge weights, using an attention mechanism that blends temporal and spatial cues. It is evaluated on four spatio-temporal datasets (PedalMe, England Covid, WindMill, ChickenPox), outperforming a range of baselines on MSE, RMSE, and MAE. The work demonstrates improved predictive accuracy and interpretability for temporal graph data, and provides public code to support reproducibility.

Abstract

Graph neural networks (GNN) have shown significant capabilities in handling structured data, yet their application to dynamic, temporal data remains limited. This paper presents a new type of graph attention network, called TempoKGAT, which combines time-decaying weight and a selective neighbor aggregation mechanism on the spatial domain, which helps uncover latent patterns in the graph data. In this approach, a top-k neighbor selection based on the edge weights is introduced to represent the evolving features of the graph data. We evaluated the performance of our TempoKGAT on multiple datasets from the traffic, energy, and health sectors involving spatio-temporal data. We compared the performance of our approach to several state-of-the-art methods found in the literature on several open-source datasets. Our method shows superior accuracy on all datasets. These results indicate that TempoKGAT builds on existing methodologies to optimize prediction accuracy and provide new insights into model interpretation in temporal contexts.

TempoKGAT: A Novel Graph Attention Network Approach for Temporal Graph Analysis

TL;DR

TempoKGAT addresses forecasting on temporal graphs by integrating time-decaying weights with selective top-k neighbor aggregation into a Graph Attention Network. The model applies temporal decay to node features and restricts neighbor aggregation to the most informative neighbors based on edge weights, using an attention mechanism that blends temporal and spatial cues. It is evaluated on four spatio-temporal datasets (PedalMe, England Covid, WindMill, ChickenPox), outperforming a range of baselines on MSE, RMSE, and MAE. The work demonstrates improved predictive accuracy and interpretability for temporal graph data, and provides public code to support reproducibility.

Abstract

Graph neural networks (GNN) have shown significant capabilities in handling structured data, yet their application to dynamic, temporal data remains limited. This paper presents a new type of graph attention network, called TempoKGAT, which combines time-decaying weight and a selective neighbor aggregation mechanism on the spatial domain, which helps uncover latent patterns in the graph data. In this approach, a top-k neighbor selection based on the edge weights is introduced to represent the evolving features of the graph data. We evaluated the performance of our TempoKGAT on multiple datasets from the traffic, energy, and health sectors involving spatio-temporal data. We compared the performance of our approach to several state-of-the-art methods found in the literature on several open-source datasets. Our method shows superior accuracy on all datasets. These results indicate that TempoKGAT builds on existing methodologies to optimize prediction accuracy and provide new insights into model interpretation in temporal contexts.
Paper Structure (13 sections, 11 equations, 3 figures, 2 tables, 1 algorithm)

This paper contains 13 sections, 11 equations, 3 figures, 2 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Proposed TempoKGAT Model
  4. Nodes Features Transformation with Temporal Decay
  5. Top-k Neighbor Selection Based on Edge Weights
  6. Attention Mechanism and Weighted Aggregation
  7. Experimental Setup
  8. Dataset
  9. Baseline Models
  10. Evaluation Metrics
  11. Results and Discussion
  12. Conclusion and Future Directions
  13. Acknowledgments.

Figures (3)

  • Figure 1: Example of a spatio-temporal graph's complexity, capturing evolving connections over time and space.
  • Figure 2: Diagrammatic Representation of the TempoKGAT Layer Workflow: This diagram traces the sequence of operations within the TempoKGAT layer, starting with the initial node features. It initiates with the Temporal Block, where a time decay function is applied to nodes to prioritize recent data. The process advances to the Spatial Block, depicting the selection of top-k neighbors based on edge weights, indicated by green lines, while excluded edges are in red. Following this, the focus shifts to computing attention coefficients, which gauge the significance of each chosen neighbor. The workflow concludes with an attention-weighted feature update, integrating temporal and spatial insights to refine node features for the next timestep ($t_{n+1}$).
  • Figure 3: Impact of the $k$ parameter in the top $k$ selection on the RMSE across different datasets. Each curve represents a unique dataset, showcasing how variations in $k$ in percentage, where for each dataset the minimum $k$ is 1 and the maximum $k$ is the average node degree of the dataset, affect the predictive accuracy of the model. This analysis illustrates the significance of choosing an optimal $k$ value for enhancing TempoKGAT performance across diverse graph-based tasks.