In this paper we propose two constructions of dependent normalized random measures, a class of nonparametric priors over dependent probability measures. Our constructions, which we call mixed normalized random measures (MNRM) and thinned normalized random measures (TNRM), involve (respectively) weighting and thinning parts of a shared underlying Poisson process before combining them together. We show that both MNRM and TNRM are marginally normalized random measures, resulting in well understood theoretical properties. We develop marginal and slice samplers for both models, the latter necessary for inference in TNRM. In time-varying topic modeling experiments, both models exhibit superior performance over related dependent models such as the hierarchical Dirichlet process and the spatial normalized Gamma process.
|Title of host publication||Proceedings of the 30th International Conference on Machine Learning (ICML 2013)|
|Subtitle of host publication||June 16 – June 21, 2013, Atlanta, Georgia, USA|
|Editors||Sanjoy Dasgupta, David McAllester|
|Publisher||International Machine Learning Society (IMLS)|
|Number of pages||9|
|Publication status||Published - 2013|
|Event||International Conference on Machine Learning 2013 - Atlanta, United States of America|
Duration: 16 Jun 2013 → 21 Jun 2013
Conference number: 30th
|Conference||International Conference on Machine Learning 2013|
|Abbreviated title||ICML 2013|
|Country||United States of America|
|Period||16/06/13 → 21/06/13|
Chen, C., Rao, V., Buntine, W., & Teh, Y. W. (2013). Dependent normalized random measures. In S. Dasgupta, & D. McAllester (Eds.), Proceedings of the 30th International Conference on Machine Learning (ICML 2013): June 16 – June 21, 2013, Atlanta, Georgia, USA (pp. 2006-2014). International Machine Learning Society (IMLS).