Random Jungle: A Nature-Inspired Improvement of Tree-Based Classifiers

Tree-based ensemble methods remain central to structured and tabular classification tasks; however, classical Random Forest and boosting approaches exhibit limitations in modeling complex feature interactions, handling noisy or imbalanced data, and balancing robustness with computational cost. This paper proposes Random Jungle (RJ), a heterogeneous ensemble framework that enhances structural diversity through operator-driven feature transformations while preserving the interpretability of tree-based models. RJ integrates three key components: 1) Label Density Weight (LDW) for refined probability estimation under class imbalance and noise, 2) heterogeneous operator forests based on additive, subtractive, multiplicative, and power transformations to capture complementary feature interactions, and 3) adaptive aggregation via weighted integration to balance predictive stability and representation diversity. Unlike boosting methods that rely on sequential error correction, RJ increases structural expressiveness through parallel operator diversification, offering a principled compute–accuracy trade-off. Extensive evaluation on 19 benchmark datasets against nine widely used classifiers demonstrates consistent top-ranked performance across multiple metrics. RJ achieved near-perfect accuracy on Glass Identification (99.12%), Breast Cancer Wisconsin (98.88%), and Balance Scale (96.76%), and obtained the best overall average ranking (1.28), outperforming Decision Tree (2.71) and Random Forest/XGBoost (3.28). Runtime and memory analyses further quantify the dimensional expansion and computational footprint, enabling practical assessment for real-world deployment. These results establish RJ as a robust, interpretable, and structurally diversified ensemble framework for high-dimensional classification tasks.