Social Influence Maximization for Public Health Campaigns

Qianyi Zhan, Wei Zhuo, Yuan LiuIEEE Access 2019 · DOI

Abstract

Effective public health communication relies on identifying influential individuals to disseminate critical information. Traditional Influence Maximization (IM) methods often struggle with the computational complexity of large networks and the neglect of heterogeneous user attributes. This paper introduces NE-IM (Network Embedding for Influence Maximization), a deep learning framework that combines structural network topology with rich node features into low-dimensional embeddings. By transforming the seed selection process into a latent space similarity problem, NE-IM achieves superior information spread and significant computational speedups compared to traditional simulation-based heuristics.

Introduction

In public health, the goal of a campaign is often to spread life-saving information—such as vaccination benefits or smoking cessation—as widely as possible. This is a classic Influence Maximization (IM) problem: given a social network, how do you pick a small set of "seed" users to start a chain reaction of information sharing? However, existing solutions face two primary challenges: the Heterogeneity Gap, where models ignore user-specific roles (e.g., medical professionals vs. patients), and the Efficiency Gap, where finding optimal seeds in massive networks is computationally prohibitive due to reliance on Monte Carlo simulations.

Methodology

The NE-IM framework addresses these gaps by shifting from simulation-based selection to representation learning. The core idea is to map every user into a low-dimensional vector (an embedding) that captures their "essence" in the network.

Architecture of the NE-IM framework, integrating structure and feature embeddings.

As shown in the architecture above, the framework utilizes two parallel paths:

  • Structure-Based Embedding: Uses a deep autoencoder to capture both first-order proximity (direct links) and second-order proximity (shared neighborhoods).
  • Feature-Based Embedding: Uses a Multi-Layer Perceptron (MLP) to process heterogeneous node attributes, including discrete data (gender, medical department) and continuous data (age, profile descriptions).

Instead of running thousands of "what-if" spread simulations, NE-IM calculates a relationship score matrix using the inner product of these learned vectors. The influence of a node is determined by its average relationship score with all other nodes in the latent space. This turns a massive simulation problem into a much faster linear algebra problem.

Experimental Results

The model was evaluated on four real-world public health datasets crawled from Twitter and Weibo, covering topics like smoking, hypertension, vaccination, and nutrition. These datasets ranged from sparse networks (Hypertension) to dense ones (Vaccination).

Key Findings

  • Superior Spread: In all four datasets, NE-IM outperformed baselines like CC-Heuristic, Degree Discount, and Harvester. As the number of allowed seed nodes increased, the gap between NE-IM and the baselines widened, showing better scalability.
  • Massive Speedup: Computational efficiency was a standout result. For the Vaccination dataset, NE-IM selected seeds in 6.93 seconds, while the Degree Discount baseline took 27.43 seconds.
  • Performance in Sparsity: Even in the sparse Hypertension dataset, the feature-based embedding allowed NE-IM to find influential nodes that topology-only models missed, proving that identity matters in health networks.

Conclusion and Future Work

NE-IM demonstrates that integrating identity and network structure through deep learning significantly enhances influence maximization for public health. By treating influence as a distance problem in a latent vector space, the authors bypass the primary bottlenecks of the last decade of IM research.

Future work will address the limitations of static snapshots by exploring dynamic network evolution, where the influence of a "seed" may change over time. Additionally, researchers aim to test the framework's performance under diverse diffusion models beyond the Independent Cascade (IC) model.

Summary updated

BibTeX
@article{zhan2019social,
  author = {Qianyi Zhan and Wei Zhuo and Yuan Liu},
  title = {{Social Influence Maximization for Public Health Campaigns}},
  journal = {IEEE Access},
  year = {2019},
  doi = {10.1109/access.2019.2946391},
}
Figure