OP3-RF: Sparse Oblique Random Forests for Interpretable and Efficient Ensemble Learning

We introduce OP3-RF, a sparse oblique random forest that uses the OP3 oblique decision tree as its base learner. OP3-RF embeds low-dimensional oblique splits within a random forest, constraining each hyperplane to at most three attributes so that interpretability is enforced directly during split construction rather than recovered post-hoc. Unlike conventional oblique forests, which depend on costly optimisation or yield opaque, high-dimensional models, this design preserves transparency while remaining computationally efficient. We evaluate OP3-RF across multiple benchmark datasets against standard random forests, oblique decision trees, and existing oblique forest variants. OP3-RF attains competitive classification accuracy and ROC-AUC while producing sparser, more interpretable decision structures, and it reaches near-optimal performance with comparatively few trees, indicating efficient ensemble behaviour. These results show that sparse oblique ensembles can jointly balance interpretability, computational efficiency, and predictive performance, advancing the goals of explainable and trustworthy machine learning.