Distance-Aware Learning for Inductive Link Prediction on Temporal Networks.

Area of Science:

• Computer Science
• Network Science
• Machine Learning

Background:

• Inductive link prediction on temporal networks forecasts future connections for unseen nodes.
• Current methods often rely on node attributes or common neighbors, which are not always available or reliable, especially in sparse networks.

Purpose of the Study:

• To propose a novel distance-aware learning (DEAL) approach for inductive link prediction on temporal networks.
• To address limitations of existing methods that depend on node attributes or struggle with sparse temporal network structures.

Main Methods:

• DEAL employs an adaptive sampling method to extract temporal adaptive walks, enhancing the inclusion of common neighbors.
• A dual-channel distance measuring component assesses distances in both embedding space and dynamic graph structure.
• The approach is evaluated on MathOverflow, AskUbuntu, and StackOverflow datasets.

Main Results:

• DEAL demonstrates superior performance over state-of-the-art baselines in inductive link prediction.
• Improvements in accuracy, Area Under the ROC Curve (AUC), and Average Precision (AP) were observed.
• The method shows particular effectiveness in scenarios with limited data.

Conclusions:

• The proposed DEAL approach offers a robust solution for inductive link prediction in temporal networks.
• Its ability to handle sparse data and lack of attributes makes it suitable for real-world applications.
• DEAL advances the field by providing more accurate and reliable link prediction in dynamic network environments.