Investigating social influence using embedding-based similarity and spatial homophily on the location based social networks

Abstract

Location-based social networks (LBSNs) are important platforms for business applications due to the rich combination of user relationships and geographical information. One of the key challenges in this context is identifying influential users by estimating social influence, which plays a central role in targeted advertising and behavior prediction. Prior studies suggest that user similarity is a significant factor contributing to social influence. Traditional methods typically compute similarity using structural metrics, while more recent approaches adopt embedding techniques to capture user proximity in vector space. However, these methods often reflect only limited aspects of user characteristics, overlooking the synergy between network structure and spatial behavior. To address this, our research introduces and evaluates two progressive methodological contributions. First, we propose the NSH (Node Embedding & Spatial Homophily) framework, a robust and interpretable hybrid method that linearly combines embedding-based structural similarity with spatial homophily. This unsupervised fusion provides a strong baseline for analyzing the relationship between similarity and influence. Building upon this, our core innovation is the GeoSocial2Vec framework, a novel, end-to-end unified embedding model. It learns a single representation for each user by intrinsically merging social ties and spatial co-visitation patterns within a heterogeneous graph via an adaptive random walk mechanism, advancing the state-of-the-art in joint socio-spatial representation learning. We evaluate the relationship between user similarity and social influence through Pearson correlation and regression analysis on geographically filtered segments of three real-world LBSNs: Gowalla, Brightkite, and Foursquare. Our results demonstrate that both proposed methods—where spatial homophily is integrated with similarity measures—produce a significantly stronger correlation with observed influence patterns compared to classical approaches. The GeoSocial2Vec framework, in particular, achieves superior performance, highlighting the critical role of intrinsically coupled socio-spatial learning. Furthermore, while GeoSocial2Vec offers the highest accuracy, the NSH framework maintains advantages in interpretability and computational simplicity, offering a practical tool for various application scenarios.