A0182
Title: Learning mixed latent forest models
Authors: Can Zhou - Nanjing Audit University (China) [presenting]
Xiaofei Wang - Northeast Normal University (China)
Jianhua Guo - Beijing Technology and Business University (China)
Abstract: A latent forest model is adopted for mining mixed data. Compared to the traditional latent tree model, the adopted model is more flexible with several trees, and observed variables are allowed to be internal nodes of trees. This model can capture more complex potential mechanisms behind data. The latent structural learning and the parameter estimation for this model are addressed. For structural learning, a consistent bottom-up algorithm is designed, and a theoretical guarantee is provided on a finite sample size bound for the exact structural recovery. For parameter estimation, a moment estimator algorithm is suggested, and the estimator is proven asymptotically normal. The simulation studies indicate that the algorithms performed well in learning the mixed latent forest model. Moreover, the learned mixed latent forest model had a better classification performance than the Naive Bayes model. The real data analysis shows that the learned model captured the hierarchical structure and latent information behind the Changchun mayor hotline data.
