[Paper Review] Evolving Influence Maximization.

Xudong Wu, Luoyi Fu+2 more|arXiv (Cornell University)|Apr 3, 2018

Complex Network Analysis Techniques2 references3 citations

TL;DR

This paper introduces EIM, a bandit-based framework for Influence Maximization in evolving social networks where user relationships and influence probabilities change over time. By iteratively selecting seed users and learning evolving network dynamics, EIM outperforms static IM methods in spreading awareness under temporal network evolution.

ABSTRACT

Influence Maximization (IM) aims to maximize the number of people that become aware of a product by finding the `best' set of `seed' users to initiate the product advertisement. Unlike prior arts on static social networks containing fixed number of users, we undertake the first study of IM in more realistic evolving networks with temporally growing topology. The task of evolving IM ({\bfseries EIM}), however, is far more challenging over static cases in the sense that seed selection should consider its impact on future users and the probabilities that users influence one another also evolve over time. We address the challenges through $\mathbb{EIM}$, a newly proposed bandit-based framework that alternates between seed nodes selection and knowledge (i.e., nodes' growing speed and evolving influences) learning during network evolution. Remarkably, $\mathbb{EIM}$ involves three novel components to handle the uncertainties brought by evolution:

Motivation & Objective

• Address the challenge of influence maximization in dynamic social networks where user connections and influence probabilities evolve over time.
• Overcome the limitations of static IM methods that assume fixed network structures and unchanging influence probabilities.
• Develop a framework that adaptively selects seed users while simultaneously learning evolving network dynamics.
• Handle the uncertainty introduced by temporal changes in network topology and influence propagation.
• Enable scalable and effective influence spreading in realistic, time-evolving social networks.

Proposed method

• Propose EIM, a bandit-based framework that alternates between seed selection and knowledge learning during network evolution.
• Introduce a novel exploration-exploitation strategy to balance selecting high-impact seed nodes and learning unknown influence dynamics.
• Model evolving influence probabilities as time-dependent parameters that are updated based on observed propagation outcomes.
• Use a multi-armed bandit approach to estimate the influence potential of nodes under changing network conditions.
• Incorporate node growth speed as a learnable parameter to predict future influence spread.
• Dynamically re-evaluate seed sets as new nodes and edges emerge, ensuring long-term influence maximization.

Experimental results

Research questions

• RQ1 How can influence maximization be effectively extended to evolving social networks with time-varying topologies?
• RQ2 What mechanisms enable adaptive seed selection in the presence of evolving influence probabilities and network growth?
• RQ3 How does the integration of online learning improve influence spread compared to static IM approaches?
• RQ4 To what extent can a bandit-based framework handle uncertainty in evolving network dynamics?
• RQ5 What is the impact of learning node growth speed and evolving influence on long-term influence propagation?

Key findings

• EIM significantly outperforms static IM baselines in influence spread on evolving networks, demonstrating superior adaptability.
• The bandit-based learning component enables effective exploration of uncertain influence dynamics while maintaining high exploitation of known influential nodes.
• Incorporating node growth speed as a learnable parameter improves prediction accuracy of future influence propagation.
• The framework achieves stable performance across multiple evolving network scenarios, showing robustness to temporal changes.
• EIM reduces the performance gap between static and dynamic IM settings by effectively modeling time-evolving influence probabilities.
• Empirical evaluation confirms that the iterative learning and selection process leads to higher cumulative influence spread than non-adaptive methods.