Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

HyperEvent: A Strong Baseline for Dynamic Link Prediction via Relative Structural Encoding

Jian Gao, Jianshe Wu, JingYi Ding

TL;DR

This work addresses the need for a strong yet simple baseline for dynamic link prediction in continuous-time dynamic graphs. It proposes HyperEvent, which uses handcrafted relative structural encoding of recent events and a lightweight Transformer to assess the authenticity of hyper-events, enabling scalable, parallelizable training via segmented event streams. Across five real-world temporal graph datasets, HyperEvent delivers competitive MRR performance with significantly lower complexity and training time than many complex baselines, especially on dense, temporally structured data; it underperforms primarily on highly unpredictable, sparse-context datasets. The findings suggest that carefully designed simple structural priors can yield strong baselines and provide a clear reference point for evaluating future advances in temporal graph representation learning.

Abstract

Learning representations for continuous-time dynamic graphs is critical for dynamic link prediction. While recent methods have become increasingly complex, the field lacks a strong and informative baseline to reliably gauge progress. This paper proposes HyperEvent, a simple approach that captures relative structural patterns in event sequences through an intuitive encoding mechanism. As a straightforward baseline, HyperEvent leverages relative structural encoding to identify meaningful event sequences without complex parameterization. By combining these interpretable features with a lightweight transformer classifier, HyperEvent reframes link prediction as event structure recognition. Despite its simplicity, HyperEvent achieves competitive results across multiple benchmarks, often matching the performance of more complex models. This work demonstrates that effective modeling can be achieved through simple structural encoding, providing a clear reference point for evaluating future advancements.

HyperEvent: A Strong Baseline for Dynamic Link Prediction via Relative Structural Encoding

TL;DR

This work addresses the need for a strong yet simple baseline for dynamic link prediction in continuous-time dynamic graphs. It proposes HyperEvent, which uses handcrafted relative structural encoding of recent events and a lightweight Transformer to assess the authenticity of hyper-events, enabling scalable, parallelizable training via segmented event streams. Across five real-world temporal graph datasets, HyperEvent delivers competitive MRR performance with significantly lower complexity and training time than many complex baselines, especially on dense, temporally structured data; it underperforms primarily on highly unpredictable, sparse-context datasets. The findings suggest that carefully designed simple structural priors can yield strong baselines and provide a clear reference point for evaluating future advances in temporal graph representation learning.

Abstract

Learning representations for continuous-time dynamic graphs is critical for dynamic link prediction. While recent methods have become increasingly complex, the field lacks a strong and informative baseline to reliably gauge progress. This paper proposes HyperEvent, a simple approach that captures relative structural patterns in event sequences through an intuitive encoding mechanism. As a straightforward baseline, HyperEvent leverages relative structural encoding to identify meaningful event sequences without complex parameterization. By combining these interpretable features with a lightweight transformer classifier, HyperEvent reframes link prediction as event structure recognition. Despite its simplicity, HyperEvent achieves competitive results across multiple benchmarks, often matching the performance of more complex models. This work demonstrates that effective modeling can be achieved through simple structural encoding, providing a clear reference point for evaluating future advancements.

Paper Structure

This paper contains 28 sections, 19 equations, 7 figures, 6 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Method
  4. Preliminaries and Problem Formulation
  5. Existing Frameworks
  6. Node-centric methods
  7. Event-centric Methods
  8. HyperEvent––A Simple Structural Encoding Baseline
  9. Real-time Adjacency Table Maintenance ($\textcircled{1}$ in Figure \ref{['fig_2']})
  10. Context Extraction ($\textcircled{2}$ in Figure \ref{['fig_2']})
  11. Relative Structural Encoding ($\textcircled{3}$ in Figure \ref{['fig_2']})
  12. Classifier
  13. Efficient Training Algorithm
  14. Complexity Analysis
  15. experiment
  16. ...and 13 more sections

Figures (7)

  • Figure 1: An example of a hyper-event (shopping trip) comprising five sequential transaction events (taxi, food, grocery, cinema, drink) along a timeline ($t_1$–$t_5$).
  • Figure 2: The HyperEvent prediction pipeline: 1)Temporal neighborhood interactions are maintained in real-time adjacency tables; 2) Hyper-events are extracted from source/destination node neighborhoods; 3)Structural-temporal correlations are computed as relational matrices; 4)A discriminator verifies hyper-event authenticity to predict query event occurrence.
  • Figure 3: Visualization of an example of the event correlation vector computation in the HyperEvent.
  • Figure 4: Parallelized training framework. Adjacency tables precomputed at key timestamps ($t_{1}, t_{2}, t_{3}$) anchor segmented event streams (Event Stream 1, Event Stream 2, Event Stream 3).
  • Figure 5: GPU memory usage (MB, right axis) and per-epoch training duration (seconds, left axis) of HyperEvent across varying segment counts ($n_{\text{segment}}$).
  • ...and 2 more figures