Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

On the Power of Heuristics in Temporal Graphs

Filip Cornell, Oleg Smirnov, Gabriela Zarzar Gandler, Lele Cao

TL;DR

This work shows that temporal graphs exhibit strong recency and popularity signals that can be captured with simple memorization-based heuristics, often rivaling or surpassing neural models under standard evaluation. It introduces four heuristics—Global Recency, Local Recency, Global Popularity, and Local Popularity—and a principled way to combine them, along with efficient ranking via Fenwick Trees to enable full rankings in $\mathcal{O}(\log S)$ time. A new Complementary Normalized Rank metric provides a lens for analyzing dataset predictability and the impact of temporal patterns, while extensive experiments on the Temporal Graph Benchmark and BenchTemp demonstrate robust performance and highlight evaluation protocol effects. The findings underscore the value of refined evaluation and interpretability for temporal graph models and point to future work on de-biasing and hybrid approaches that blend heuristics with neural methods.

Abstract

Dynamic graph datasets often exhibit strong temporal patterns, such as recency, which prioritizes recent interactions, and popularity, which favors frequently occurring nodes. We demonstrate that simple heuristics leveraging only these patterns can perform on par or outperform state-of-the-art neural network models under standard evaluation protocols. To further explore these dynamics, we introduce metrics that quantify the impact of recency and popularity across datasets. Our experiments on BenchTemp and the Temporal Graph Benchmark show that our approaches achieve state-of-the-art performance across all datasets in the latter and secure top ranks on multiple datasets in the former. These results emphasize the importance of refined evaluation schemes to enable fair comparisons and promote the development of more robust temporal graph models. Additionally, they reveal that current deep learning methods often struggle to capture the key patterns underlying predictions in real-world temporal graphs. For reproducibility, we have made our code publicly available.

On the Power of Heuristics in Temporal Graphs

TL;DR

This work shows that temporal graphs exhibit strong recency and popularity signals that can be captured with simple memorization-based heuristics, often rivaling or surpassing neural models under standard evaluation. It introduces four heuristics—Global Recency, Local Recency, Global Popularity, and Local Popularity—and a principled way to combine them, along with efficient ranking via Fenwick Trees to enable full rankings in time. A new Complementary Normalized Rank metric provides a lens for analyzing dataset predictability and the impact of temporal patterns, while extensive experiments on the Temporal Graph Benchmark and BenchTemp demonstrate robust performance and highlight evaluation protocol effects. The findings underscore the value of refined evaluation and interpretability for temporal graph models and point to future work on de-biasing and hybrid approaches that blend heuristics with neural methods.

Abstract

Dynamic graph datasets often exhibit strong temporal patterns, such as recency, which prioritizes recent interactions, and popularity, which favors frequently occurring nodes. We demonstrate that simple heuristics leveraging only these patterns can perform on par or outperform state-of-the-art neural network models under standard evaluation protocols. To further explore these dynamics, we introduce metrics that quantify the impact of recency and popularity across datasets. Our experiments on BenchTemp and the Temporal Graph Benchmark show that our approaches achieve state-of-the-art performance across all datasets in the latter and secure top ranks on multiple datasets in the former. These results emphasize the importance of refined evaluation schemes to enable fair comparisons and promote the development of more robust temporal graph models. Additionally, they reveal that current deep learning methods often struggle to capture the key patterns underlying predictions in real-world temporal graphs. For reproducibility, we have made our code publicly available.

Paper Structure

This paper contains 12 sections, 3 figures, 4 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Method
  4. The Notion of Recency
  5. The Notion of Popularity
  6. Combining Heuristics
  7. Experiments and Results
  8. Heuristics as Analysis Tools
  9. Conclusion
  10. Additional Results
  11. Efficient Implementation of Heuristics
  12. Complementary Normalized Ranking Plots

Figures (3)

  • Figure 1: Complementary Normalized Rank (CNR) plots comparing optimistic ($R^{+}$) and pessimistic ($R^{-}$) ranks across various heuristics and their combinations. Each curve shows a method's performance using the CNR metric across percentiles of all edges, illustrating its ranking effectiveness.
  • Figure 2: Complementary Normalized Ranking Plots showing optimistic ($R^{+}$) and pessimistic ($R^{-}$) ranks for different heuristics and their combinations, for TGB (panels a-f) and BenchTemp (panels g-o) datasets. Each curve represents a method's performance across different percentiles of all edges in the dataset, illustrating how well it ranks them overall.
  • Figure : Recency and Popularity Heuristics