A Generalized Inverted Dirichlet Predictive Model for Activity Recognition Using Small Training Data

Jiaxun Guo, Manar Amayri, Wentao Fan, Nizar Bouguila

Abstract: In this paper, we develop the predictive distribution of the generalized inverted Dirichlet (GID) mixture model using local variational inference. The main goal is to be able to tackle classification problems involving small training data sets. The two main ingredients of the proposed predictive model are the GID distribution which provides flexibility for the modeling of semi-bounded data that are naturally generated by different sensors outputs and the efficient of variational inference as a deterministic approximation to fully Bayesian approaches. The merits of the proposed model are shown via synthetic data and a real application that concerns activities recognition.