Regional Isolation Forest: A Novel Approach for Enhanced Anomaly Detection and Interpretability

Anomaly detection is widely used in safety- and risk sensitive applications, where both reliable detection and interpretable decision support are important. Among unsupervised methods, Isolation Forest and its variants remain attractive because of their efficiency and simple tree-based structure. However, existing hyperplane-based isolation mechanisms may produce geometry-induced score artifacts, such as ghost regions in anomaly score maps. Although geometry-aware variants of Isolation Forest have been proposed to mitigate these artifacts, residual artifacts may still arise under complex data distributions, and detector-specific interpretability is less explored, thereby leaving room for further improvement in practical applications. To address these issues, this article proposes a Regional Isolation Forest, denoted as ReIF, for improved anomaly detection and interpretation. ReIF constrains the spatial extent of recursive partitioning through bounded regional splits and controls the attenuation of inner and outer partition ranges by two geometry-related parameters. On this basis, a model-specific interpretability strategy, RIFFI, is further developed, including a local component for instance-level anomaly diagnosis and a global component for dataset-level feature summarization. Extensive experiments on simulated and real-world datasets validate the proposed framework from multiple perspectives. The results show that ReIF can effectively alleviate geometry-induced artifacts and improve detection performance under complex data distributions. Additional sensitivity analysis, shape ablation, runtime evaluation, and comparisons with SHAP and LIME further demonstrate the robustness, efficiency, and practical interpretability of the proposed method.