A comprehensive Survey of the Actual Causality Literature

The study of causality has recently gained traction in computer science. Formally capturing causal reasoning would allow computers to answer “Why”-questions and would result in significant advances in fields such as verification, machine learning, explainability, legal reasoning and algorithmic fairness. To accomplish this, one needs to be able to infer type causal relationships, i.e. general statements about causal dependencies, from data and then use those relationships to identify the actual causes of an event in a given situation; such causes are referred to as token causes. To the best of our knowledge, there does not exist a comprehensive survey, reviewing the state of the art of formal systems for token causality. The present thesis addresses this deficit. The literature review that we have performed operates on three different levels of granularity. The first considered the literature landscape itself as an object of study, employing network analysis techniques to identify important publications, authors and research communities. The second is a classical literature review, where a subset of the collected literature is investigated in detail, to extract, describe and categorise the tools used for formalising causation. This includes the languages for encoding causal relationships, the various definitions that try to capture token causality, as well as the benchmark used to test the capabilities of those definitions. In the third part we describe and compare the four main token causality definitions, w.r.t. the most prominent benchmarks in the literature. This last part also required some original work, as not all the examples are found in the literature.