Exploring Human Mobility Correlations Using Semisupervised Hypergraph Clustering

Bound by social bonds, people often travel in groups, exhibiting mutual correlations in their mobility behaviors, known as mobility correlations. However, conventional graphs, which describe pairwise connections between vertices, fail to adequately capture the complex many-to-many connection characteristics of multiple individuals. To address this limitation, we propose a novel semisupervised hypergraph clustering (SSHC) model aimed at unraveling the intricate correlations among individuals before discerning their latent clusters. Each individual is characterized by features that encompass spatial, temporal, and attributive diversity. The high-order correlations concealed among multiple individuals are captured using a hypergraph framework featuring hyperedges. Building upon this foundation, we fuse individuals’ latent feature representations, extracted via an auto-encoder, with their structural representations derived from a hypergraph convolutional network. Through this fusion, we identify clusters of similar individuals exhibiting mobility correlations. Notably, our model incorporates a semisupervised learning scheme, leveraging priori patterns gleaned from a small pool of ground-truth data to better express latent patterns within unlabeled datasets. Experimental validations conducted in five public datasets corroborate the superiority of our proposed model over baseline methods in elucidating complex correlations within vast datasets. In addition, experiments conducted on Beijing Transit Smart Card Data showcase promising results in identifying correlated individuals with similar mobility patterns. The delineation of groups of individuals exhibiting mobility correlations is of paramount importance, particularly in accurately detecting their connection networks during public emergencies, such as the transmission of dangerous pathogens.