Link prediction in complex networks via matrix perturbation and decomposition

Area of Science:

• Network Science
• Data Mining
• Computational Social Science

Background:

• Link prediction in complex networks is crucial for understanding network structure and function.
• Existing methods often struggle with incomplete data and noise, limiting prediction accuracy.
• Previous research focused on either data completion or noise reduction individually.

Purpose of the Study:

• To develop a unified link prediction method that integrates data completion and noise reduction.
• To enhance the accuracy of predicting missing links in complex networks.
• To extend the method's applicability to weighted directed networks.

Main Methods:

• A novel link prediction method combining data completion via structural generalization and noise reduction via decomposition techniques.
• Exploiting global topological information from the adjacency matrix.
• Generalizing the perturbation method by extracting the symmetric part of the adjacency matrix.

Main Results:

• The proposed method significantly improved prediction accuracy compared to individual approaches.
• Experimental results on real-world networks demonstrated superior performance over existing methods and local indices.
• The method was successfully extended to handle weighted directed networks.

Conclusions:

• Combining data completion and noise reduction offers a more robust approach to link prediction.
• The novel method provides a significant advancement in predicting missing links in complex networks.
• This approach has broad applicability across various scientific domains involving network analysis.