The paper starts by providing background information on dynamic graphs and their applications, as well as the representation and learning of dynamic graphs. It then introduces a novel taxonomy to categorize the 81 DGNN models covered in the survey. The models are classified into four groups for discrete-time dynamic graphs (DTDG) and seven groups for continuous-time dynamic graphs (CTDG), based on their structural features, use of methods, and dynamic modeling techniques.
Next, the paper presents a detailed overview of 12 existing DGNN training frameworks, including 5 DTDG frameworks and 7 CTDG frameworks. It discusses the key features, supported functionalities, and optimization strategies of these frameworks.
The paper then introduces commonly used evaluation benchmarks for DGNN models, covering 20 diverse graph datasets across various application domains, such as social networks, interaction networks, event networks, trade networks, and traffic networks. It also provides an overview of the commonly used evaluation metrics for DGNN models, including binary classification performance, link prediction, and training efficiency.
To provide a comprehensive comparison of DGNN models and frameworks, the paper conducts experiments on six standard graph datasets, evaluating nine representative DGNN models and three DGNN frameworks. The evaluation focuses on convergence accuracy, training efficiency, and GPU memory usage, enabling a thorough comparison of performance across different models and frameworks.
Finally, the paper analyzes the key challenges in the DGNN field and suggests potential research directions for future work, such as enhancing model expressiveness, improving training efficiency and scalability, and addressing emerging application demands.
To Another Language
from source content
arxiv.org
Key Insights Distilled From
by ZhengZhao Fe... at arxiv.org 05-02-2024
https://arxiv.org/pdf/2405.00476.pdfDeeper Inquiries