A multi-granularity perspective for spatial profiling of mobile apps

Abstract The prevalence of mobile apps has greatly changed people’s lives, generating myriads of data and creating new research opportunities. By exploiting mobile apps’ data, existing studies have primarily investigated app users’ usage patterns. Few studies have focused on profiling the geospatial distributions of individual apps’ usage patterns on a large scale. A major challenge for profiling app usage is the heterogeneity and sparsity of individual apps’ data. In this study, we propose a multi-level mixture of kernel density estimation (mlKDE) model for robust profiling of the geospatial distribution of any given app. A major advantage of this model is that it can leverage aggregate information from groups of related apps, and adaptively train the weights for different groups. Using a real-world, large-scale app usage dataset covering more than one thousand apps available on the market, we demonstrate that our model can effectively characterize the distributions of app usage in the real world, and has considerable advantages when compared to the baseline models reported in the literature.