Optimizing the Amount of Models Taken into Consideration During Model Selection in Bayesian Networks

Graphical model selection from data embodies several difficulties. Among them, it is specially challenging the size of the sample space of models on which one should carry out model selection, even considering only a modest amount of variables. This becomes more severe when one works on those graphical models where some variables may be responses to other. This is the case of Bayesian Networks that are modeled by acyclic digraphs. In this paper we try to reduce the amount of models taken into consideration during model selection. The less amount of models considered, the less amount of steps performed to end the model selection process, and therefore the less computational effort required to fit data and models, We propose a simple idea: to select models from sample spaces of lower dimensionand use them as starting models for sample spaces of an upper dimension. Plots on experimental results are provided on four different synthetic datasets. They show that the main idea reduces substantially the steps taken during model selection, in comparison to a greedy model selection procedure.