Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Input Snapshots Fusion for Scalable Discrete-Time Dynamic Graph Neural Networks

QingGuo Qi, Hongyang Chen, Minhao Cheng, Han Liu

TL;DR

The paper tackles scalability in discrete-time dynamic graphs by merging multiple input snapshots into a single temporal graph and integrating Hawkes processes with GNNs to capture time-decayed influence across edges. It introduces Input Snapshots Fusion (ISF) and Hawkes-GNN variants (Hawkes-GCN, Hawkes-GAT) anchored by a Hawkes excitation matrix, enabling efficient full-batch and mini-batch training for future link prediction. The authors provide theoretical connections to graph denoising with time decay, extensive experiments on eight datasets showing superior performance and substantial memory savings, and ablations demonstrating the efficacy of temporal-decay message passing. The approach offers a scalable, robust framework for modeling temporal edges in DTDGs with broad applicability to large-scale dynamic graph tasks while avoiding heavy sequential encoders.

Abstract

In recent years, there has been a surge in research on dynamic graph representation learning, primarily focusing on modeling the evolution of temporal-spatial patterns in real-world applications. However, within the domain of discrete-time dynamic graphs, the exploration of temporal edges remains underexplored. Existing approaches often rely on additional sequential models to capture dynamics, leading to high computational and memory costs, particularly for large-scale graphs. To address this limitation, we propose the Input {\bf S}napshots {\bf F}usion based {\bf Dy}namic {\bf G}raph Neural Network (SFDyG), which combines Hawkes processes with graph neural networks to capture temporal and structural patterns in dynamic graphs effectively. By fusing multiple snapshots into a single temporal graph, SFDyG decouples computational complexity from the number of snapshots, enabling efficient full-batch and mini-batch training. Experimental evaluations on eight diverse dynamic graph datasets for future link prediction tasks demonstrate that SFDyG consistently outperforms existing methods.

Input Snapshots Fusion for Scalable Discrete-Time Dynamic Graph Neural Networks

TL;DR

The paper tackles scalability in discrete-time dynamic graphs by merging multiple input snapshots into a single temporal graph and integrating Hawkes processes with GNNs to capture time-decayed influence across edges. It introduces Input Snapshots Fusion (ISF) and Hawkes-GNN variants (Hawkes-GCN, Hawkes-GAT) anchored by a Hawkes excitation matrix, enabling efficient full-batch and mini-batch training for future link prediction. The authors provide theoretical connections to graph denoising with time decay, extensive experiments on eight datasets showing superior performance and substantial memory savings, and ablations demonstrating the efficacy of temporal-decay message passing. The approach offers a scalable, robust framework for modeling temporal edges in DTDGs with broad applicability to large-scale dynamic graph tasks while avoiding heavy sequential encoders.

Abstract

In recent years, there has been a surge in research on dynamic graph representation learning, primarily focusing on modeling the evolution of temporal-spatial patterns in real-world applications. However, within the domain of discrete-time dynamic graphs, the exploration of temporal edges remains underexplored. Existing approaches often rely on additional sequential models to capture dynamics, leading to high computational and memory costs, particularly for large-scale graphs. To address this limitation, we propose the Input {\bf S}napshots {\bf F}usion based {\bf Dy}namic {\bf G}raph Neural Network (SFDyG), which combines Hawkes processes with graph neural networks to capture temporal and structural patterns in dynamic graphs effectively. By fusing multiple snapshots into a single temporal graph, SFDyG decouples computational complexity from the number of snapshots, enabling efficient full-batch and mini-batch training. Experimental evaluations on eight diverse dynamic graph datasets for future link prediction tasks demonstrate that SFDyG consistently outperforms existing methods.
Paper Structure (45 sections, 1 theorem, 20 equations, 6 figures, 5 tables, 1 algorithm)

This paper contains 45 sections, 1 theorem, 20 equations, 6 figures, 5 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Preliminaries
  3. Link Prediction Tasks in DTDGs
  4. Graph Convolutional Networks
  5. Graph Attention Networks
  6. GNNs as Solving Graph Denoising Problem
  7. Hawkes Process
  8. The SFDyG Model
  9. Input Snapshots Fusion
  10. Hawkes Processes Based GNN
  11. Hawkes-GCN
  12. Hawkes-GAT
  13. Loss Function
  14. Minibatch Training Algorithm
  15. Complexity Analysis
  16. ...and 30 more sections

Key Result

Theorem 3.1

When we adopt the Laplacian matrix ${{\bf L}} = {{\bf D}} - {\mathcal{C}}$, where ${\bf D} = diag(\sum_j{{\mathcal{C}}_{ij}})$, the graph denoising problem (Equaition eq:denoising) was extended to temporal domain with the assumption that the smoothness of edges decays over time.

Figures (6)

  • Figure 1: Training setting for the task of link prediction in dynamic graphs.
  • Figure 2: The GPU memory usage of SFDyG and representative baselines on six datasets.
  • Figure 3: The comparison of average epoch time and maximum GPU memory usage during mini-batch training. The dotted line represents one dataset, while the endpoints illustrate the execution metrics on that dataset.
  • Figure 4: Performance of Hawkes-GAT with varying numbers of negative samples during training.
  • Figure 5: The distribution of MRR@100 on the UCI dataset with different negative samples for training.
  • ...and 1 more figures

Theorems & Definitions (2)

  • Definition 3.1: Hawkes excitation matrix
  • Theorem 3.1: Hawkes temporal graph denoising