Data usually present in heterogeneous sources. When dealing with multiple data sources, existing models often treat them independently and thus can not explicitly model the correlation structures among data sources. To address this problem, we propose a full Bayesian nonparametric approach to model correlation structures among multiple and heterogeneous datasets. The proposed framework, first, induces mixture distribution over primary data source using hierarchical Dirichlet processes (HDP). Once conditioned on each atom (group) discovered in previous step, context data sources are mutually independent and each is generated from hierarchical Dirichlet processes. In each specific application, which covariates constitute content or context(s) is determined by the nature of data. We also derive the efficient inference and exploit the conditional independence structure to propose (conditional) parallel Gibbs sampling scheme. We demonstrate our model to address the problem of latent activities discovery in pervasive computing using mobile data. We show the advantage of utilizing multiple data sources in terms of exploratory analysis as well as quantitative clustering performance.
展开▼
