Temporal influence maximization via continuous-time graph neural networks and deep reinforcement learning.

Area of Science:

• Network Science
• Artificial Intelligence
• Computational Social Science

Background:

• Traditional influence maximization (IM) methods fail on dynamic networks due to static assumptions.
• Real-world social systems exhibit continuous evolution and bursty interactions, invalidating static network models.
• Existing temporal IM methods discretize time, losing fine-grained dependencies and failing to model non-stationary patterns.

Purpose of the Study:

• To develop a novel framework, TempRL-IM, for influence maximization in dynamic networks.
• To address the limitations of static and discretized temporal IM approaches.
• To leverage continuous-time dynamics for more accurate and efficient influence spread prediction.

Main Methods:

• Integration of Continuous-Time Graph Neural Networks (CTGNNs) for temporal dependency encoding.
• Utilization of a Double Deep Q-Network (DDQN) agent for optimal seed selection.
• Development of a temporal reinforcement learning framework (TempRL-IM) for dynamic network analysis.

Main Results:

• TempRL-IM achieves 15-28% higher influence spread compared to state-of-the-art methods.
• The framework demonstrates 3-10x faster inference speeds.
• Strong transferability across networks with similar temporal characteristics was observed.

Conclusions:

• TempRL-IM effectively models continuous-time dynamics in social networks without discretization artifacts.
• The proposed framework offers significant improvements in both accuracy and efficiency for influence maximization.
• TempRL-IM shows promise for large-scale applications like viral marketing and epidemic containment.