Bayesian predictive modelling: application to aircraft short-term conflict alert system

Bayesian Model Averaging (BMA), computationally feasible using Markov Chain Monte Carlo (MCMC), is a well-known method for reliable estimation of predictive distributions. The use of decision tree (DT) models for the averaging enables experts not only to estimate a predictive posterior but also to interpret models of interest and estimate the importance of predictor factors that are assumed to contribute to the prediction. The MCMC method generates parameters of DT models in order to explore their posterior distributions and to draw samples from the models. However, these samples can often over-represent DT models of an excessive size, which in cases of real-world applications affects the results of BMA. When this happens, it is unlikely for a DT model that provides Maximum a Posteriori probability to explain the observed data with high accuracy. We propose a new technology in order to estimate and interpret predictive posteriors. In our experiments with aircraft short-term conflict alerts, we show how this technology can be used for analysing uncertainties in detections of conflicts.