Tree-Weighting

Tree-Weighting applies weighting strategies to Random Forests and individual decision trees to improve cross-study predictive performance and robustness in multi-study learning.

Key Features:

• Multi-Study Learning Framework: Classifiers are independently trained on multiple studies and ensemble weights are assigned based on cross-study prediction ability.
• Random Forest-based Learners: Uses Random Forests as the single-study learner and implements two weighting strategies: weighting entire forests and extracting then weighting individual trees.
• Tree-Level Weighting and Layered Ensembling: Assigns weights at the individual-tree level across multiple ensembling layers to increase robustness of the final ensemble.
• Analysis of Tree Structure and Weighting Benefits: Investigates correlations between ensemble weights and tree structures to identify when direct tree weighting outperforms forest-level weighting.

Scientific Applications:

• Genomic multi-study predictive modeling: Applied to genomic datasets to improve predictive accuracy and robustness across independent studies, supporting analyses of complex biological processes and disease mechanisms.

Methodology:

Independent Random Forest classifiers are trained on each study; cross-study prediction ability is computed to derive ensemble weights; two weighting strategies are implemented (forest-level weighting and extraction/weighting of individual trees); weights may be assigned at the tree level forming multiple layers of ensembling; analysis of correlations between ensemble weights and tree structures is performed.