SUMMARY

Understanding the causal relations governing sociotechnical systems allows designers to better understand system dynamics, identify the root cause of issues, and subsequently design more sustainable systems. Visual representations help with this understanding and can be achieved with a multitude of approaches. Bayesian Networks (BNs) represented by Directed Acyclic Graphs (DAGs) are a particularly promising approach because they compactly and intuitively convey a lot information about dependencies and independencies in a system. BNs are often generated manually by experts, but a large number of data driven BN learning algorithms have been developed that may aid non-experts in making decisions. Most BN learning algorithms work best with data that satisfies the Causal Sufficiency Assumption, Markov Assumption, and DAG Faithfulness Assumption; all of which are not commonly satisfied in real world socio-technical data. This research aims to answer the question “What is the relationship between the performance of different BN learning algorithms and the validity of different assumptions about a dataset?” The performance of BN learning algorithms is characterized by their learned structure’s match with the true causal network and ability to perform predictive inference. By comparing datasets where the three assumptions are and are not satisfied this research can learn how algorithms perform with and without those assumptions. Better understanding this relation will allow designers of socio-technical systems to select the best BN learning algorithms and data preprocessing approaches for their problems. The research found that the learning algorithms decreased in their ability to perform predictive inference and recover the expert specified structure as the number of valid data assumptions decreased. The research also found a strong case for utilizing an additive Bayesian network to learn structure and perform accurate inference.