Event-Aware Prompt Learning for Dynamic Graphs

Basic Information

• Paper ID: 2510.11339
• Title: Event-Aware Prompt Learning for Dynamic Graphs
• Authors: Xingtong Yu¹, Ruijuan Liang², Xinming Zhang², Yuan Fang¹
• Institutions: ¹Singapore Management University, ²University of Science and Technology of China
• Classification: cs.LG cs.AI
• Publication Date: October 13, 2025 (arXiv preprint)
• Paper Link: https://arxiv.org/abs/2510.11339

Abstract

Real-world graphs typically evolve through a series of events, modeling dynamic interactions between objects across different domains. In dynamic graph learning, Dynamic Graph Neural Networks (DGNNs) have become the mainstream solution. Recently, prompt learning methods have been explored on dynamic graphs. However, existing methods primarily focus on capturing node-time relationships while neglecting the influence of historical events. This paper proposes EVP, an event-aware prompt learning framework for dynamic graphs that can serve as a plugin to enhance existing methods' ability to leverage historical event knowledge. First, a series of historical events are extracted for each node, and an event adaptation mechanism is introduced to align fine-grained features of these events with downstream tasks. Second, an event aggregation mechanism is proposed to effectively integrate historical knowledge into node representations. Finally, extensive experiments on four public datasets are conducted to evaluate and analyze EVP.

Research Background and Motivation

Problem Definition

The core problem addressed by this research is: how to effectively leverage historical event knowledge in dynamic graph learning to improve downstream task performance. Specifically, it includes:

1. Inconsistency between pretraining and downstream task objectives: Traditional DGNNs are typically trained on link prediction tasks, while downstream tasks may be node classification, leading to objective mismatch.
2. Neglect of historical event influence: Existing dynamic graph prompt learning methods primarily focus on node-time relationships, ignoring historical event knowledge from graph evolution processes.

Significance

1. Real-world relevance: Real-world graph structures evolve through event sequences, such as users posting on Reddit or creating pages on Wikipedia.
2. Performance improvement: Historical events contain valuable knowledge that can predict future events and enhance node temporal representation learning.
3. Parameter efficiency: Prompt learning only requires adjusting lightweight prompts, particularly effective in low-resource scenarios.

Limitations of Existing Methods

1. Traditional DGNNs: Suffer from objective inconsistency between pretraining and downstream tasks.
2. Dynamic graph pretraining methods: Face similar knowledge transfer difficulties due to objective mismatch.
3. Existing dynamic graph prompt learning: Methods such as DyGPrompt and TIGPrompt only consider node-time interactions, ignoring structural evolution information from historical events.

Core Contributions

1. Proposes EVP framework: The first event-aware prompt learning framework for dynamic graphs that can serve as a plugin to enhance existing dynamic graph learning methods.
2. Event adaptation mechanism: Designs event prompts to capture fine-grained features of historical events, adapting to downstream task requirements.
3. Event aggregation mechanism: Combines dynamic prompts and temporal decay functions to effectively integrate historical event knowledge.
4. Comprehensive experimental validation: Validates EVP's effectiveness and generalizability as a plugin across four benchmark datasets.

Methodology Details

Task Definition

• Input: Dynamic graph G = (V, E, T), where V is the node set, E is the edge set, and T is the temporal axis.
• Event definition: Each edge (vᵢ, vⱼ, t) ∈ E represents an interaction event from node vᵢ to vⱼ at time t.
• Objective: Leverage historical event knowledge to improve performance on temporal link prediction and node classification tasks.

Model Architecture

1. Event Extraction

For node v at time t, extract K historical events:

Ev,t = {E¹v,t, E²v,t, ..., EᴷV,t}

Each event is represented as:

Eᵏv,t = (v, uᵏv,t, zᵏv,t)

where zᵏv,t is the timestamp of the k-th event and uᵏv,t is the interacting object node.

2. Event Adaptation Mechanism

First, compute event embeddings:

eᵏv,t = FUSE(hv, huᵏv,t)

Then perform adaptation through event prompts:

êᵏv,t = pe ⊙ eᵏv,t

where pe is a learnable event prompt vector and ⊙ denotes element-wise multiplication.

3. Event Aggregation Mechanism

Combine temporal decay functions and dynamic prompts:

ẽv,t = Σᵏ₌₁ᴷ pᵏdy · êᵏv,t

where pdy ∈ Rᴷ is the dynamic prompt and pᵏdy is its k-th dimension.

Simultaneously consider temporal decay:

êv,t = Σᵏ₌₁ᴷ exp(t - zᵏv,t) · êᵏv,t

4. Prompt Tuning

Integrate historical event embeddings with node embeddings:

ĥv,t = hv,t + ẽv,t

Technical Innovations

1. First introduction of event awareness: Unlike existing methods, EVP explicitly models the influence of historical events on current behavior.
2. Dual-layer aggregation strategy: Combines temporal decay (reflecting temporal proximity) and dynamic prompts (capturing pattern similarity).
3. Plug-and-play design: Seamlessly integrates into DGNNs, pretraining methods, and prompt learning approaches.
4. Fine-grained event modeling: Adapts fine-grained event features to specific downstream tasks through event prompts.

Experimental Setup

Datasets

Four benchmark datasets are used:

Evaluation Metrics

• Link prediction: AUC-ROC
• Node classification: AUC-ROC
• Support for both transductive and inductive settings.

Baseline Methods

Four categories of baseline methods:

1. Traditional DGNNs: ROLAND, TGAT, TGN, TREND, GraphMixer
2. Dynamic graph pretraining: DDGCL, CPDG
3. Static graph prompt learning: GraphPrompt, ProG
4. Dynamic graph prompt learning: TIGPrompt, DyGPrompt

Implementation Details

• Data split: 80% pretraining, 20% downstream tasks (1%/1%/18% for training/validation/testing)
• Low-resource setting: Only 30 events used for downstream tuning
• 100 repeated samplings to construct different tasks, with 5 runs per task and averaging results.

Experimental Results

Main Results

EVP achieves the best performance across all tasks:

Temporal Link Prediction (Transductive):

• Wikipedia: 98.47±0.80% (vs DyGPrompt 94.33±0.12%)
• Reddit: 99.85±0.14% (vs DyGPrompt 96.82±0.06%)
• MOOC: 98.16±0.54% (vs DyGPrompt 70.17±0.75%)
• Genre: 99.90±0.02% (vs DyGPrompt 87.02±1.63%)

Node Classification:

• Wikipedia: 87.18±3.21% (vs DyGPrompt 82.09±6.43%)
• Reddit: 76.77±7.93% (vs DyGPrompt 74.00±3.10%)
• MOOC: 78.78±4.04% (vs DyGPrompt 77.78±5.08%)

Plugin Effect Verification

EVP as a plugin significantly improves performance across 7 different methods:

Example with TGAT:

• Transductive link prediction: Improved from 55.78% to 76.50% (Wikipedia)
• Inductive link prediction: Improved from 48.21% to 76.65% (Wikipedia)
• Node classification: Improved from 67.00% to 79.03% (Wikipedia)

Ablation Study

Comparison of three EVP variants:

• EVP-EP: Uses only event prompts with direct summation aggregation
• EVP-DP: Uses only dynamic prompts without temporal decay
• EVP-TD: Uses only temporal decay without dynamic prompts

Results show that complete EVP outperforms all variants across all settings, validating the necessity of each component.

Hyperparameter Sensitivity Analysis

Analysis of historical event count K shows:

• Link prediction: Optimal performance at K=9
• Node classification: Optimal performance at K=3
• Excessive historical events may introduce noise and degrade performance.

Related Work

Dynamic Graph Learning

1. DGNNs: Update node embeddings by dynamically aggregating neighbor information with temporal factors.
2. Dynamic graph pretraining: Adopts "pretrain-finetune" paradigm but suffers from objective mismatch.
3. Temporal representation learning: Designs various strategies to capture graph structure changes.

Dynamic Graph Prompt Learning

Existing methods such as DyGPrompt and TIGPrompt design time-aware and node-aware prompts but neglect historical event influence. EVP fills this gap.

Conclusions and Discussion

Main Conclusions

1. Historical event knowledge is crucial: EVP significantly improves dynamic graph learning performance by leveraging historical events.
2. Strong generalizability: As a plugin, it can enhance multiple existing methods.
3. Reasonable design: Both event adaptation and aggregation mechanisms contribute importantly to performance improvement.

Limitations

1. Computational complexity: Requires extracting and processing K historical events for each node.
2. Hyperparameter sensitivity: K value requires tuning for different tasks.
3. Limited event definition: Currently only considers interaction events between nodes.
4. Insufficient interpretability: Lacks in-depth analysis of learned event patterns.

Future Directions

1. More complex event modeling: Consider event types, intensity, and other attributes.
2. Adaptive event selection: Dynamically determine the optimal number of historical events for each node.
3. Theoretical analysis: Provide theoretical guarantees for EVP's effectiveness.
4. Large-scale applications: Validate on larger real-world dynamic graphs.

In-Depth Evaluation

Strengths

1. Strong novelty: First to introduce event awareness in dynamic graph prompt learning, filling an important gap.
2. Clever design: Event adaptation and aggregation mechanisms are well-designed, balancing temporality and pattern characteristics.
3. Comprehensive experiments: Validated across multiple datasets and tasks with detailed ablation studies and plugin effect analysis.
4. High practical value: Plug-and-play design makes it easy to apply to existing methods.
5. Significant performance improvements: Achieves substantial performance gains across all test scenarios.

Weaknesses

1. Weak theoretical foundation: Lacks theoretical analysis of why historical events are effective.
2. Insufficient computational overhead analysis: Detailed analysis of added computational and storage costs is missing.
3. Simplified event modeling: Models events simply as node pair interactions, potentially losing important information.
4. Dataset limitations: Primarily validated on relatively small datasets, lacking large-scale validation.
5. Insufficient interpretability: Lacks analysis and visualization of learned historical event patterns.

Impact

1. Academic contribution: Provides new research directions and insights for the dynamic graph learning field.
2. Practical value: Plug-and-play characteristics make it easily adoptable by the community.
3. Inspirational value: May inspire more research on event modeling in graph evolution processes.
4. Reproducibility: Provides detailed implementation details and hyperparameter settings.

Applicable Scenarios

1. Social network analysis: User behavior prediction, community evolution analysis.
2. Recommendation systems: Dynamic recommendations based on historical interactions.
3. Knowledge graphs: Temporal evolution modeling of entity relationships.
4. Bioinformatics: Dynamic analysis of protein interaction networks.
5. Financial risk management: Anomaly detection in transaction networks.

References

The paper cites important works from dynamic graph learning, graph neural networks, and prompt learning fields, providing readers with comprehensive background knowledge. Key references include classical dynamic graph methods such as TGN and TGAT, as well as prompt learning methods like GraphPrompt and DyGPrompt.

Overall Assessment: This is a high-quality research paper that makes important contributions to the field of dynamic graph prompt learning. The EVP framework is well-designed, experiments are comprehensive, and it has excellent practical value and academic significance. Despite some theoretical and experimental limitations, it represents significant progress in this field.