Social Influence Maximization for Public Health Campaigns

With the rapid growth of online medical treatments, social media has become a prevalent platform to spread health-related information. For public health campaigns, how to diffuse the message broadly and efficiently in the social network is a classical problem, which is named Influence Maximization (IM). IM can be considered as an algorithmic problem of finding a small set of network users as seed nodes that maximizes the spread of influence for a piece of information under a certain influence cascade model. With a lot of researches focusing on this field, most existing solutions are proposed to extract seed nodes only considering pure topological structures. In practice, nodes in health-related networks are often abundantly accompanied with other types of meaningful information such as node attributes (identity), second-order proximity, and edge attributes (user relationship), while conventional IM models overlook these information. Therefore, a general framework for incorporating the heterogeneous information into an IM model could be potentially helpful to find underlying seed nodes in health-related networks. Moreover, the stability and the computational cost are also the main challenges that current IM models face. To bridge these gaps, in this paper, we propose NE-IM (Network Embedding for Influence Maximization), a method that aspires to address both problems using representation learning. NE-IM composes of two components: Structure-based embedding and feature-based embedding. They are the projections of network structures and heterogeneous information respectively in a low dimensional space, so that each node in health-related network can be represented as a fix dimensional vector. Experimental results show that our methods significantly outperform baseline approaches.