Robust Temporal Link Prediction in Dynamic Complex Networks via Stable Gated Models With Reinforcement Learning

Area of Science:

• Complex Networks
• Machine Learning
• Data Mining

Background:

• Temporal link prediction is crucial for understanding time-varying networks but faces challenges with adversarial attacks and diverse evolutionary patterns.
• Existing methods struggle to maintain robustness and adapt to the dynamic nature of complex networks.

Purpose of the Study:

• To propose a novel, robust temporal link prediction architecture named SAGE-RL.
• To enhance adaptability to distinct network evolutionary patterns and defend against adversarial attacks.

Main Methods:

• Developed a SAGE-RL architecture comprising a State Encoding Network (SEN) and a Self-Adaptive Policy Network (SPN).
• Introduced a novel stable gate within SEN to ensure spatiotemporal dependencies and defend against attacks.
• Utilized SPN to adapt SEN to various evolutionary patterns by approximating optimal action functions.

Main Results:

• SAGE-RL demonstrated superior performance over state-of-the-art methods in temporal link prediction precision and stability across five real-world benchmarks.
• The architecture proved effective in defending against various adversarial attacks.
• Successfully applied temporal link prediction to shipping transaction networks, forecasting potential transaction risks.

Conclusions:

• SAGE-RL offers a robust and adaptive solution for temporal link prediction in dynamic complex networks.
• The proposed stable gate and self-adaptive policy network significantly improve resilience and accuracy.
• The framework has practical implications for risk forecasting in transaction networks.