Modeling continuous distributions in hybrid Bayesian networks using mixtures of polynomials with tails

A new approach to modeling continuous distributions in hybrid Bayesian networks (BNs) is presented. It is based on Mixtures of Polynomials (MoPs) with tails, named as tMoPs. This proposal is a variation of the usual MoP model, now including tails and several other improvements in the learning process. The adequate modeling of tails in variable distributions is relevant theoretically and for many reals applications, in which rare phenomena may have a great impact. The proposed approach has been designed to exploit the flexibility of the tMoP model to fit different continuous data distributions. This is especially relevant in those distributions with zones of density close to zero, in which polynomial fitting may be difficult. In these situations, tMoPs allow a polynomial fit in parts with higher density and the use of tails in areas with lower density. This permits a better global fit, without loss of overall accuracy and yielding a relatively simple density function. Learning algorithms for tMoPs conditional probability distributions with up to two parents of any type are developed. These tMoPs may be integrated into hybrid Bayesian networks to represent conditional probability distributions, thus allowing to perform probabilistic reasoning, such as causal inference, sensitivity analysis, and other decision-making operations. The suitability of tMoPs is evaluated in several ways, using a large set of real datasets with data of different natures. The experiments include: the analysis of goodness-of-fit with several continuous and pseudo-continuous variables, the optimization of certain parameters and the effect of variable selection and graph structure when using tMoPs in BNs, and finally the evaluation of the predictive ability of hybrid BNs based on tMoPs in classification and regression. Results show the good behavior of our proposal, with the tMoP hybrid Bayesian networks being equally accurate or outperforming other techniques in most scenarios, in addition to providing a more informative and convenient probabilistic model.