Deep graph representation learning for influence maximization with accelerated inference.

Area of Science:

• Computational Social Science
• Network Science
• Machine Learning

Background:

• Influence maximization (IM) aims to select seed users to maximize influence spread in social networks.
• Classical IM approaches are nearing theoretical performance limits.
• Learning-based IM methods offer better generalization but face challenges in efficiency, diffusion pattern characterization, and adaptability.

Purpose of the Study:

• To develop a novel framework, DeepIM, to address limitations in current learning-based influence maximization methods.
• To enable data-driven, end-to-end learning of latent seed set representations and diverse diffusion patterns.
• To infer optimal seed sets under flexible node-centrality-based budget constraints.

Main Methods:

• Designed the DeepIM framework for generative characterization of seed set latent representations.
• Integrated learning of diversified information diffusion patterns in a data-driven, end-to-end manner.
• Developed a novel objective function for seed set inference under node-centrality constraints.

Main Results:

• DeepIM effectively characterizes latent representations and learns diffusion patterns.
• The novel objective function allows for flexible seed set selection under budget constraints.
• Extensive analyses on synthetic and real-world datasets demonstrate DeepIM's superior performance.

Conclusions:

• DeepIM offers a robust and adaptable solution for influence maximization problems.
• The framework overcomes key challenges, paving the way for more effective network influence strategies.
• DeepIM shows significant potential for applications in targeted marketing and information dissemination.