Measuring the robustness of link prediction algorithms under noisy environment

Area of Science:

• Network Science
• Data Science
• Computational Social Science

Background:

• Link prediction in complex networks estimates future node interactions.
• Current validation often assumes noise-free network data.
• Real-world networks frequently contain inaccuracies like missing or fake links.

Purpose of the Study:

• To comprehensively study the robustness of existing link prediction algorithms.
• To understand the impact of network data inaccuracies on prediction accuracy.
• To propose a metric for quantifying the robustness of link prediction methods.

Main Methods:

• Simulated introduction of missing, fake, and swapped links in network data.
• Evaluation of twenty-two link prediction algorithms under these noisy conditions.
• Development of an index to quantify algorithm robustness.

Main Results:

• Missing links were found to be more detrimental to prediction accuracy than fake or swapped links.
• A novel index was proposed to quantify the robustness of link prediction methods.
• Some algorithms, despite lower initial accuracy, demonstrated reliable performance in noisy network conditions.

Conclusions:

• The robustness of link prediction algorithms in real-world, noisy networks is critical.
• Network data inaccuracies, particularly missing links, pose a significant challenge.
• Certain link prediction methods offer more reliable performance in the presence of network noise.