Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Node-Time Conditional Prompt Learning In Dynamic Graphs

Xingtong Yu, Zhenghao Liu, Xinming Zhang, Yuan Fang

TL;DR

DYGPROMPT addresses the misalignment between pre-training on temporal link prediction and downstream tasks in dynamic graphs by introducing dual prompts (node and time) and dual condition-nets that generate time- and node-aware prompts. The approach couples a pre-training phase with temporal link prediction and a lightweight downstream-tuning phase that freezes the backbone and learns prompts and condition-nets, enabling efficient adaptation in data-scarce settings. Empirical results across four benchmarks show superior performance over state-of-the-art DGNNs, graph pre-training, static prompts, and existing dynamic prompts, with ablations confirming the importance of both prompts and their conditional generation. The framework demonstrates robustness across backbones and scales to large datasets, offering a practical route to improving downstream performance in evolving graphs.

Abstract

Dynamic graphs capture evolving interactions between entities, such as in social networks, online learning platforms, and crowdsourcing projects. For dynamic graph modeling, dynamic graph neural networks (DGNNs) have emerged as a mainstream technique. However, they are generally pre-trained on the link prediction task, leaving a significant gap from the objectives of downstream tasks such as node classification. To bridge the gap, prompt-based learning has gained traction on graphs, but most existing efforts focus on static graphs, neglecting the evolution of dynamic graphs. In this paper, we propose DYGPROMPT, a novel pre-training and prompt learning framework for dynamic graph modeling. First, we design dual prompts to address the gap in both task objectives and temporal variations across pre-training and downstream tasks. Second, we recognize that node and time features mutually characterize each other, and propose dual condition-nets to model the evolving node-time patterns in downstream tasks. Finally, we thoroughly evaluate and analyze DYGPROMPT through extensive experiments on four public datasets.

Node-Time Conditional Prompt Learning In Dynamic Graphs

TL;DR

DYGPROMPT addresses the misalignment between pre-training on temporal link prediction and downstream tasks in dynamic graphs by introducing dual prompts (node and time) and dual condition-nets that generate time- and node-aware prompts. The approach couples a pre-training phase with temporal link prediction and a lightweight downstream-tuning phase that freezes the backbone and learns prompts and condition-nets, enabling efficient adaptation in data-scarce settings. Empirical results across four benchmarks show superior performance over state-of-the-art DGNNs, graph pre-training, static prompts, and existing dynamic prompts, with ablations confirming the importance of both prompts and their conditional generation. The framework demonstrates robustness across backbones and scales to large datasets, offering a practical route to improving downstream performance in evolving graphs.

Abstract

Dynamic graphs capture evolving interactions between entities, such as in social networks, online learning platforms, and crowdsourcing projects. For dynamic graph modeling, dynamic graph neural networks (DGNNs) have emerged as a mainstream technique. However, they are generally pre-trained on the link prediction task, leaving a significant gap from the objectives of downstream tasks such as node classification. To bridge the gap, prompt-based learning has gained traction on graphs, but most existing efforts focus on static graphs, neglecting the evolution of dynamic graphs. In this paper, we propose DYGPROMPT, a novel pre-training and prompt learning framework for dynamic graph modeling. First, we design dual prompts to address the gap in both task objectives and temporal variations across pre-training and downstream tasks. Second, we recognize that node and time features mutually characterize each other, and propose dual condition-nets to model the evolving node-time patterns in downstream tasks. Finally, we thoroughly evaluate and analyze DYGPROMPT through extensive experiments on four public datasets.
Paper Structure (26 sections, 11 equations, 4 figures, 9 tables, 1 algorithm)

This paper contains 26 sections, 11 equations, 4 figures, 9 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Preliminaries
  4. Proposed Approach
  5. Overall framework
  6. Pre-training
  7. Dual prompts
  8. Dual condition-nets
  9. Prompt tuning
  10. Experiments
  11. Experimental setup
  12. Performance comparison with baselines
  13. Ablation studies
  14. Impact of backbones and hyperparameters
  15. Conclusions
  16. ...and 11 more sections

Figures (4)

  • Figure 1: Motivation of DyGPrompt. (a) Users comment on different topics over time. (b) Evolving node-time patterns in Wikipedia, as node and time mutually characterize each other.
  • Figure 2: Overall framework of DyGPrompt.
  • Figure 3: Visualization of output embedding space of nodes.
  • Figure 4: Sensitivity of $\alpha$.