An integer programming model for designing causal networks

Abstract

We propose a novel mixed integer programming formulation for the design of causal discovery networks. The model takes a set of rules that indicate sta-tistical dependency relations between features of a given dataset, the so-called d-connection and d-separation relations, and aims to fit a casual network with minimum (weighted) violations. Allowing feedback cycles and latent confounders, our formulation stands out from most of the existing attempts in the literature. Although our model can work as an unsupervised machine learning model, it possesses the necessary flexibility for the decision-maker to enter known causal relations. The performance of our model is tested with several synthetic datasets.