Occupancy estimation in smart buildings using predictive modeling in imbalanced domains

Jiaxun Guo, Manar Amayri, Fatma Najar, Wentao Fan, Nizar Bouguila

Abstract: This paper introduces a novel approach for occupancy estimation in smart buildings. In particular, we focus on the challenging yet common situation where the amount of training data is small and imbalanced (i.e. the classes are not approximately equally represented). Our model is based on two parts namely predictive modelling, performed via the inverted Dirichlet mixture (IDMM), and an over-sampling approach that we propose. The first part, in which the main goal is to tackle the small training data problem, concerns the calculation of the predictive distribution of the IDMM by marginalizing over its parameters, with their posterior distributions, which are estimated by a Bayesian variational inference approach that we develop. Based on over-sampling, the second part can be viewed as a complement to tackling the imbalanced domains problem. Extensive experiments and simulations involving synthetic data as well as real data extracted from smart buildings sensors show the merits of our statistical framework.