Authors:
Walisson Ferreira
1
;
Mark Song
2
and
Luis Zarate
2
Affiliations:
1
Centro Universitário UNA, Brazil
;
2
Pontificia Universidade Católica de Minas Gerais, Brazil
Keyword(s):
Causal Inference, Formal Concept Analysis, FCA, Markov Equivalence, Causal Bayesian Networks, Causal Relationship, Bayesian Networks, Attributes Implication.
Abstract:
One of the problems during the construction of Causal Bayesian Network based on constraint algorithms occurs when it is not possible to orient edges between nodes due to Markov Equivalence. In this scenario this article presents the use of Formal Concept Analysis (FCA), specially attributes implication, as an alternative to support the definition of the direction of the edges. To do this it was applied algorithms of Bayesian learners (PC) and FCA in a data set containing 12 attributes and 5,473 records of surgeries performed in Belo Horizonte - Brazil. According to the results, although attribute implication did not necessarily mean causality, the implication rules were useful in defining edges orientation on the Bayesian network learned by PC Algorithm. The results of FCA were validated through intervention using do-calculus and by an expert in the domain. Therefore, as result of this paper, it is presented a heuristic to direct edges between nodes when the direction is unknown.