Influence Function Learning in Information Diffusion Networks.

Area of Science:

• Social Network Analysis
• Information Diffusion Modeling
• Machine Learning

Background:

• Traditional methods for determining user influence in social networks rely on a two-stage process: first modeling information diffusion, then calculating influence.
• The accuracy of these traditional methods is heavily dependent on the correctness of the diffusion model, which is challenging to validate with real-world data.

Purpose of the Study:

• To develop a novel approach for learning user influence directly from information diffusion cascade data.
• To bypass the need for pre-specifying a particular diffusion model, thereby improving robustness and efficiency.

Main Methods:

• Proposed a new parameterization for influence functions, viewing them as coverage functions represented by a convex combination of random basis functions.
• Developed an efficient maximum likelihood-based algorithm to learn these influence functions directly from cascade data.

Main Results:

• The proposed method can provably learn influence functions with low sample complexity.
• The approach demonstrates robustness against unknown diffusion models.
• Empirical results show significant performance improvements over existing methods on both synthetic and real-world datasets.

Conclusions:

• Directly learning influence functions from cascade data is a viable and effective alternative to traditional two-stage approaches.
• The novel parameterization and learning algorithm offer a more robust, efficient, and accurate way to assess user influence in social networks.