IM2Vec: Representation learning-based preference maximization in geo-social networks

Abstract

Recent advancements in mobile technology have facilitated location-based social networks. The location-based influence maximization problem, which aims to find top influential seed users for promoting a target location to attract the most individuals, has drawn increasing attention. However, the existing studies largely neglect the importance of user preference, which considerably hinders their practicability. In addition, time efficiency is a critical issue for handling large-scale datasets. To address the above problems, we propose a new framework named IM2Vec, which incorporates representation learning into location-based influence maximization problem. Specifically, we first propose a representation learning model, All2Vec, to capture user preferences for the target location from check-in records, which takes both user preference and geographical location influence into consideration. Then, based on the learned user preferences, we extend the reverse influence sampling (RIS) model and propose a highly efficient preference maximization algorithm, which ensures a (1-1/e-∊)-approximate solution with a substantially lower sample size. The experimental results of the two tasks (future visitor prediction and influence maximization) on two real geo-social networks show that the All2Vec model achieves considerably higher accuracy in future visitor prediction, and IM2Vec exhibits a higher influence spread and a lower running time than the state-of-the-art baselines.