Determinable and interpretable network representation for link prediction.

Area of Science:

• Complex Systems Science
• Network Science
• Machine Learning

Background:

• Complex networks are used to describe physical, social, and brain systems.
• Network representation maps network attributes into low-dimensional vector spaces.
• Current machine learning methods for network representation lack interpretability and determinacy.

Purpose of the Study:

• To propose novel, physically-grounded, and interpretable node representation methods.
• To develop a network-based model (AIProbS) for evaluating these methods in link prediction.
• To address limitations of black-box machine learning approaches in network analysis.

Main Methods:

• Proposed two deterministic and interpretable node representation methods based on a physical perspective.
• Developed the Adaptive and Interpretable ProbS (AIProbS) model for link prediction using node representations.
• Evaluated AIProbS on small datasets with non-unified training/test distributions.

Main Results:

• AIProbS achieved state-of-the-art precision on small datasets, outperforming baseline models.
• The proposed methods demonstrated a favorable trade-off between precision, determinacy, and interpretability compared to machine learning methods.
• Effectiveness and generalization of the node representations were validated.

Conclusions:

• The proposed methods offer determinable and interpretable node representations for complex networks.
• AIProbS provides a robust alternative for link prediction, particularly for small or challenging datasets.
• This work benefits industrial companies with limited computing resources and a need for interpretable models.