Summary of Towards Ideal Temporal Graph Neural Networks: Evaluations and Conclusions After 10,000 Gpu Hours, by Yuxin Yang et al.
Towards Ideal Temporal Graph Neural Networks: Evaluations and Conclusions after 10,000 GPU Hours
by Yuxin Yang, Hongkuan Zhou, Rajgopal Kannan, Viktor Prasanna
First submitted to arxiv on: 28 Dec 2024
Categories
- Main: Machine Learning (cs.LG)
- Secondary: Social and Information Networks (cs.SI)
GrooveSquid.com Paper Summaries
GrooveSquid.com’s goal is to make artificial intelligence research accessible by summarizing AI papers in simpler terms. Each summary below covers the same AI paper, written at different levels of difficulty. The medium difficulty and low difficulty versions are original summaries written by GrooveSquid.com, while the high difficulty version is the paper’s original abstract. Feel free to learn from the version that suits you best!
| Summary difficulty | Written by | Summary |
|---|---|---|
| High | Paper authors | High Difficulty Summary Read the original abstract here |
| Medium | GrooveSquid.com (original content) | Medium Difficulty Summary The proposed comparative evaluation framework for Temporal Graph Neural Networks (TGNNs) addresses the limitations in existing works by performing a design space search across well-known TGNN modules. The framework is based on a unified, optimized code implementation, allowing for clear accuracy comparisons and efficient runtime. This study investigates three critical questions in TGNN design: module efficiency, effectiveness correlations with dataset patterns, and interplay between multiple modules. Key findings include the outperformance of recent neighbor sampling and attention aggregator over uniform neighbor sampling and MLP-Mixer aggregator; the efficacy of static node memory as an alternative to dynamic node memory; and the importance of considering repetition patterns in datasets for choosing between static or dynamic node memory. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary Temporal Graph Neural Networks (TGNNs) are powerful tools that model dynamic interactions across various domains. Existing works on TGNN modeling often neglect the design space, leading to suboptimal designs. This study proposes a framework that compares well-known TGNN modules and evaluates their performance in a unified code implementation. The results show that recent neighbor sampling and attention aggregator perform better than other methods, while static node memory can be an effective alternative to dynamic node memory. Understanding how these modules interact with dataset patterns is crucial for designing more general and effective TGNNs. |
Keywords
» Artificial intelligence » Attention
