Use of a structure aware discretisation algorithm for Bayesian networks applied to water quality predictions

Bayesian networks have become a popular modelling technique in many fields, however there are several design decisions that, if poorly made, can result in models with insufficient evidence to make good predictions. One such decision is how to discretise the continuous nodes. The lack of a commonly accepted algorithm for achieving this makes it a difficult task for novice data modellers. We present a structure aware discretisation algorithm that minimises the number of missing values in the conditional probability tables by taking into account the network structure. It also prevents users from having to specify the exact number of bins. Results from two water quality case studies in south-east Queensland showed that the algorithm has potential to improve the discretisation process over equal case discretisation and demonstrates the suitability of Bayesian networks for this field. © 2019 International Association for Mathematics and Computers in Simulation (IMACS)