Tracking influence between naïve Bayes models using score-based structure learning
Ritesh Ajoodha, Benjamin Rosman · 2017
Current structure learning practices in Bayesian networks have been developed to learn the structure between observable variables and learning latent parameters independently. One exception establishes a variant of EM for learning the structure of Bayesian networks in the presence of incomplete data [1]. However, no method has demonstrated learning the influence structure between latent variables that describe (or are learned from) a number of observations. We present a method that learns a set of naïve Bayes models (NBMs) independently given a partitioned set of observations, and then attempts to track the high-level influence structure between every NBM. The latent parameters of each model are then relearned to fine-tune the influence distribution between models for density estimation of new observations. Experimental results are provided which demonstrate the effectiveness of our non-parametric method. Applications of this method include knowledge discovery and density estimation in situations where we do not fully observe characteristics of the environment.