Generating structure of latent variable models for nested data

Probabilistic latent variable models have been successfully used to capture intrinsic characteristics of various data. However, it is nontrivial to design appropriate models for given data because it requires both machine learning and domain-specific knowledge. In this paper, we focus on data with nested structure and propose a method to automatically generate a latent variable model for the given nested data, with the proposed method, the model structure is adjustable by its structural parameters. Our model can represent a wide class of hierarchical and sequential latent variable models including mixture models, latent Dirichlet allocation, hidden Markov models and their combinations in multiple layers of the hierarchy. Even when deeply-nested data are given, where designing a proper model is difficult even for experts, our method generate an appropriate model by extracting the essential information. We present an efficient variational inference method for our model based on dynamic programming on the given data structure. We experimentally show that our method generates correct models from artificial datasets and demonstrate that models generated by our method can extract hidden structures of blog and news article datasets.