Finding Causality in Socio-Technical Systems: A Comparison of Bayesian Network Structure Learning Algorithms

Abstract

A wide number of Bayesian Network (BN) structure learning algorithms have been developed for a variety of applications. The purpose of this research is to shed light on which of these BN structure learning algorithms work best with small, amalgamated socio-technical datasets in an attempt to better understand such systems and improve their design. BN structure learning algorithms have not been widely used for socio-technical problems due to the small and disparate natures of the data describing such systems. This research tested four widely used learning algorithms given two test cases: a simulated ALARM network data set as a baseline and a novel socio-technical network data set combining Divvy bike’s bike share data and Chicago weather data as the more challenging design case. After testing the K2, PC, Sparse Candidate Algorithm (SCA), and Min-Max Hill Climbing (MMHC) algorithm, results indicate that all of the algorithms tested are capable of giving insight into the novel dataset’s most likely causal structure given the real socio-technical data. It should be noted that the convergence with the real world socio-technical data was significantly slower than with the simulated ALARM network dataset. The conditional independence (PC) algorithm also exhibited an interesting pattern in that it diverged farther away from the novel socio-technical network’s most likely structure when given very large datasets, preferring a denser network with more edges. The resulting network structures from all of the algorithms suggest that an opportunity exists to increase ridership by women during commuting hours in the Divvy bike program.