SUME

Abstract

Recent years have witnessed a rapid proliferation of personalized mobile applications, which poses a pressing need for accurate user demographics inference. Facilitated by the prevalent smart devices, the ubiquitously collected mobility trace presents a promising opportunity to infer user demographics at large-scale. In this paper, we propose a novel Semantic-enhanced Urban Mobility Embedding (SUME) model, which learns dense representation vectors for user demographic inference by jointly modelling the physical mobility patterns and the semantic of urban mobility. Specifically, SUME models urban mobility as a heterogeneous network of users and locations, with various types of edges denoting the physical visitation and semantic similarities. Moreover, SUME optimizes the node representation vectors with two alternating objective functions that preserve the feature in physical and semantic domains, respectively. As a result, it is able to capture the effective signals in the heterogeneous urban mobility network. Empirical experiments on two real-world mobility traces show the proposed model significantly out-performs all state-of-the-art baselines with an accuracy margin of 8.6%~14.3% for occupation, gender, age, education and income inference. In addition, further experiments show SUME is able to reveal meaningful correlations between user demographics and the mobility patterns in spatial, temporal and urban structure domain.