An explainable multi-objective genetic programming approach to infer Boolean network

Abstract

Boolean networks can reflect the causal relationships between different parts of discrete complex systems and predict their state transitions, thereby providing efficient and qualitative insights into such systems. Numerous methods have been investigated to deduce Boolean networks from observed temporal data. However, existing algorithms focus on improving accuracy while neglecting inference explainability. To offer inference explainability without sacrificing accuracy, this paper proposes an explainable multi-objective genetic programming approach to infer large-scale Boolean networks. Unlike existing single-objective algorithms, a new explainable optimization objective is introduced by calculating the average mutual information of the tree nodes. Subsequently, we develop a bio-inspired operation to fully utilize elite solutions and enhance the exploration capability. Additionally, a penalty term for syntax trees is introduced to mitigate overfitting and improve explainability by limiting the tree size. Extensive experiments demonstrate that the proposed approach is interpretable and outperforms current leading algorithms in terms of accurately inferring large-scale Boolean networks.