Link Prediction in Complex Networks Using Average Centrality-Based Similarity Score

Area of Science:

• Network Science
• Data Mining
• Computational Social Science

Background:

• Link prediction is vital for understanding network evolution across diverse fields.
• Existing methods often rely on local or global centrality measures for similarity scoring.
• There is a need for improved link prediction accuracy.

Purpose of the Study:

• To propose novel link prediction methods using average centrality measures.
• To evaluate the effectiveness of these new measures against existing techniques.
• To enhance the accuracy of identifying future network connections.

Main Methods:

• Developed four novel similarity measures: Similarity based on Average Degree (SACD), Betweenness (SACB), Closeness (SACC), and Clustering Coefficient (SACCC).
• Calculated node centrality scores, averaged them across the graph, and derived similarity scores using common neighbors.
• Applied centrality scores to common neighbors, identifying nodes with above-average centrality.

Main Results:

• The proposed average centrality measures significantly outperformed existing local similarity measures.
• Achieved an average improvement of 24% in Area Under the Receiver Operating Characteristic (AUROC) and 49% in Area Under Precision-Recall (AUPR) over existing methods.
• Demonstrated a 31% AUROC and 51% AUPR improvement over recent link prediction measures.

Conclusions:

• The novel approach using average centrality measures offers superior performance for link prediction.
• These findings have implications for improving network analysis in various domains.
• The proposed methods provide a more accurate way to predict future links in complex networks.