Temporal network embedding framework with causal anonymous walks representations.

Area of Science:

• Graph Machine Learning
• Network Representation Learning
• Temporal Network Analysis

Background:

• Dynamic (temporal) network analysis presents challenges for traditional representation learning methods.
• Existing network embedding techniques are often limited to static graphs.
• Accurate node and edge encoding is crucial for graph machine learning tasks.

Purpose of the Study:

• To propose a novel approach for dynamic network representation learning.
• To develop a benchmark pipeline for evaluating temporal network embeddings.
• To provide a comprehensive comparison framework for temporal network representation learning.

Main Methods:

• Utilized Temporal Graph Networks (TGNs).
• Developed a custom message-generating function by extracting Causal Anonymous Walks.
• Established a benchmark pipeline for evaluating temporal network embeddings.

Main Results:

• The proposed model demonstrated superior performance compared to state-of-the-art baseline models.
• Evaluated performance across various transductive/inductive node and edge classification tasks.
• Showcased applicability and superior performance in a real-world credit scoring task for a European bank.

Conclusions:

• The novel approach effectively addresses dynamic network representation learning challenges.
• The benchmark pipeline facilitates comprehensive evaluation of temporal network embeddings.
• The model shows significant potential for real-world applications in graph machine learning.