Nonparametric Clustering with Distance Dependent Hierarchies

Abstract

The distance dependent Chinese restaurant pro-cess (ddCRP) provides a flexible framework for clustering data with temporal, spatial, or other structured dependencies. Here we model mul-tiple groups of structured data, such as pixels within frames of a video sequence, or paragraphs within documents from a text corpus. We pro-pose a hierarchical generalization of the ddCRP which clusters data within groups based on dis-tances between data items, and couples clusters across groups via distances based on aggregate properties of these local clusters. Our hddCRP model subsumes previously proposed hierarchi-cal extensions to the ddCRP, and allows more flexibility in modeling complex data. This flexi-bility poses a challenging inference problem, and we derive a MCMC method that makes coordi-nated changes to data assignments both within and between local clusters. We demonstrate the effectiveness of our hddCRP on video segmenta-tion and discourse modeling tasks, achieving re-sults competitive with state-of-the-art methods. 1