GInGER: Genetic domain partitioning for interpretable rule extraction and induction

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Symbolic knowledge-extraction techniques aim to provide human-interpretable descriptions of predictive models through compact and expressive representations. Amongst these, approaches based on input domain partitioning offer a natural way to combine local explanations with a global view of model behaviour. However, the construction of the partitions determines both the spatial structure and the complexity of the resulting symbolic model. Furthermore, existing methods often rely on rigid partitioning strategies, require discretisation of continuous targets, or produce complex and overlapping structures that hinder interpretability. This work introduces GInGER (Genetic INterval Generation for Explainable Rules), a framework that formulates symbolic model construction as the direct, genetic optimisation of a complete, flat partitioning of the input domain according to a fitness function. Candidate solutions encode the coordinates of the partition cuts, leading to a non-overlapping hypercubic partitioning to be translated into a symbolic model. Each region can be associated with a class label, a constant value, or a local linear model, enabling a unified treatment of classification and regression without mandatory target discretisation. GInGER can be applied both as a post-hoc explainer, to approximate the behaviour of black-box models, and as an ante-hoc method to directly induce interpretable models from data. An extensive empirical evaluation on multiple classification and regression data sets shows that the proposed approach achieves a favourable trade-off between predictive performance and model compactness, often producing more concise symbolic representations than competing methods while maintaining comparable accuracy. • Extraction of compact and interpretable hypercube-based representations. • Genetic optimisation of non-overlapping hypercubic partitions of the input space. • Unified support for classification and regression without discretisation. • Competitive accuracy with reduced model complexity and enhanced interpretability.