Identifying accurate link predictors based on assortativity of complex networks.

Area of Science:

• Network Science
• Data Mining
• Computational Biology

Background:

• Link prediction (LP) identifies unknown interactions in complex networks like social and biological systems.
• Network assortativity, measuring node similarity in connections, is a key structural property.
• Existing link prediction metrics often overlook assortativity's influence.

Purpose of the Study:

• To investigate link prediction metrics based on network assortativity profiles.
• To evaluate the performance of link prediction models across varying assortativity levels.
• To identify optimal link prediction strategies for different network structures.

Main Methods:

• Generated complex networks with controlled assortativity levels.
• Employed and combined neighborhood similarity and preferential attachment link prediction models.
• Assessed link prediction accuracy using the area under the precision-recall curve.

Main Results:

• Link prediction models exploring larger neighborhoods (e.g., CH2-L2, CH2-L3) showed consistent accuracy in both assortative and disassortative networks.
• Common neighbor metrics excelled in assortative networks.
• A combination of common neighbors and node degree proved effective for disassortative networks.

Conclusions:

• The choice of link prediction model should consider the network's assortativity profile for optimal performance.
• Larger neighborhood exploration offers robust link prediction across diverse network types.
• Tailored link prediction strategies enhance the analysis of incomplete network data.