Automated Hypotheses Generation via Combinatorial Causal Optimization

A powerful form of causal inference employed in many tasks, such as medical diagnosis, criminology, root cause analysis, biology, is abduction. Given an effect, it aims at generating a plausible and useful set of explanatory hypotheses for its causes. This article formulates the abductive hypotheses generation activity as an optimization problem, introducing a new class called Combinatorial Causal Optimization Problems (CCOP). In a CCOP, solutions are in the form of cause-effect combinations: algorithms are required to construct hypothetical solutions automatically assessed for plausibility - a mechanism mimicking the human reasoning when he skims the best solutions from a set of hypotheses - and for novelty with respect to already known solutions. The paper presents the CCOP formulation and four real-world benchmark problems from various domains, released along with artefacts to implement, run and properly evaluate algorithms for CCOP solutions. Then, for illustrative purpose, four conventional evolutionary algorithms are customized to solve CCOPs. Their application demonstrates the possibility of generating useful solutions (i.e., novel and realistic hypotheses for a given effect), but also evidences a great margin for improvement in terms of ratio of good vs bad solutions.

Automated summary

The article introduces a new class called Combinatorial Causal Optimization Problems (CCOP) and discusses its application in causal inference. The research uses optimization algorithms to automatically construct cause-effect solutions for plausibility and novelty, providing useful solutions for real-world problems in various domains.