Birds of a Feather Purchase Together: Accurate Social Network Inference using Transaction Data

Social networks play a crucial role in providing valuable contextual information across disciplines and applications. However, the unobservable nature of physical-world social connections has led to the development of social network inference. Existing approaches rely on co-occurrences and universal thresholds to infer social networks from spatiotemporal data. Yet, these methods suffer from two limitations: disregarding individual social preferences and failing to address "familiar strangers". Our analysis reveals that relying solely on common spatiotemporal data is inadequate for accurate social network inference. Fortunately, the availability of extensive transaction data, encompassing spatiotemporal and consumption information, presents an opportunity. Our approach involves integrating individuals' lifestyles with co-occurrences, driven by the fact that different lifestyles entail distinct social preferences and that true friends share similar lifestyles. However, we face two significant challenges: flexible extraction of lifestyle features and personalized threshold setting. To overcome these challenges, we propose nonparametric methods applicable to various scenarios and leverage domain knowledge for threshold determination. Evaluation on a real dataset of over 2, 000 individuals demonstrates an impressive improvement of over 20% in F1-score compared to the baselines.