MorphBoost: Self-Organizing Universal Gradient Boosting with Adaptive Tree Morphing

Traditional gradient boosting algorithms employ static tree structures with fixed splitting criteria that remain unchanged throughout training, limiting their ability to adapt to evolving gradient distributions and problem-specific characteristics across different learning stages. This work introduces MorphBoost, a new gradient boosting framework featuring self-organizing tree structures that dynamically morph their splitting behavior during training. The algorithm implements adaptive split functions that evolve based on accumulated gradient statistics and iteration-dependent learning pressures, enabling automatic adjustment to problem complexity. Key innovations include: (1) morphing split criterion combining gradient-based scores with information-theoretic metrics weighted by training progress; (2) automatic problem fingerprinting for intelligent parameter configuration across binary/multiclass/regression tasks; (3) vectorized tree prediction achieving significant computational speedups; (4) interaction-aware feature importance detecting multiplicative relationships; and (5) fast-mode optimization balancing speed and accuracy. Comprehensive benchmarking across 10 diverse datasets against competitive models (XGBoost, LightGBM, GradientBoosting, HistGradientBoosting, ensemble methods) demonstrates that MorphBoost achieves state-of-the-art performance, outperforming XGBoost by 0.84% on average. MorphBoost secured the overall winner position with 4/10 dataset wins (40% win rate) and 6/30 top-3 finishes (20%), while maintaining the lowest variance (σ=0.0948) and highest minimum accuracy across all models, revealing superior consistency and robustness. Performance analysis across difficulty levels shows competitive results on easy datasets while achieving notable improvements on advanced problems due to higher adaptation levels.