Using a Model of Human Cognition of Causality to Orient Arcs in Structural Learning of Bayesian Networks

Abstract

In this thesis, I present three novel heuristic algorithms for learning the structure of Bayesian networks (BNs). Two of the algorithms are based on Constructing an Undirected Graph Using Markov Blankets (CrUMB), and differ in the way they orient arcs. CrUMB- uses traditional arc orientation and CrUMB+ uses a model of human cognition of causality to orient arcs. The other algorithm, SC*, is based on the Sparse Candidate (SC) algorithm. I compare the average qualitative and quantitative performances of these algorithms with two state-of-the-art algorithms, PC and Three Phase Dependency Analysis (TPDA) algorithms. There are correctness proofs for both these algorithms, and both are implemented in software packages. The average performance of these algorithms is evaluated using one-way, within-group Analysis of Variance (ANOVA). I also apply BN structure learning to a real world dataset of drug-abuse patients who are also criminal justice offenders. The purpose of this application is to address two key issues: 1) does drug treatment increase technical violations and arrests/incarceration, which in turn influences probation, and 2) does drug treatment lead to more probation, which in turn influences violations and arrests/ incarceration? The BN models learned on this dataset were validated using k-fold cross-validation. The key contributions of this thesis are 1) the development of novel algorithms to address some of the disadvantages of existing approaches including the use of a model of human cognition of causation to orient arcs, and 2) the application of BN structure learning to a dataset coming from a domain where research and analysis have been limited to traditional statistical methods.