Weakening the faithfulness assumption in causal discovery

Thesis (PhD(Computer and Information Science))--University of South Australia, 2019. Includes bibliographical references (pages [94]-99) Learning causal relationships from data has been an active research problem as causal information helps to predict the effect of manipulations. A widely adopted framework is using graphs to represent causal structures and relate the causal structures to probability distributions via various assumptions. Two well-known assumptions are the causal Markov assumption and the causal Faithfulness assumption. While the former is almost universally accepted by the practicing researchers on causal discovery, the latter is often regarded as questionable. This is because, exact violations of faithfulness could be part of the design in some systems. Even if the Faithfulness is not exactly violated, with finite data, “almost violations” of faithfulness can occur due to errors in statistical tests of conditional independence. Such almost violations of Faithfulness can bring serious challenges to causal discovery and cannot be reasonably assumed away. In this thesis, we focus on discovering causal relationships from observational data and we are primarily concerned with the causal Faithfulness assumption.