LAVASET

LAVASET integrates latent variable analysis into a stochastic ensemble of decision trees to account for feature correlations and improve feature-importance estimation and predictive performance on high-dimensional correlated biological datasets.

Key Features:

• Latent Variable Derivation: LAVASET derives latent variables from the distance characteristics of each feature to capture and incorporate correlation factors during tree splitting.
• Enhanced Feature Importance Determination: By considering the local neighborhood of features, LAVASET distributes importance across correlated feature neighborhoods and reduces false-positive importance assigned to single noisy features.
• Robustness to Noisy Features: LAVASET mitigates the influence of individual noisy features by focusing on the collective influence of feature clusters, improving reliability relative to traditional Random Forests (RFs).
• Performance Across Diverse Data Types: LAVASET has been evaluated on simulated and real 1D datasets and on high-dimensional 3D datatypes, yielding prediction accuracies that are largely non-inferior to those of traditional RFs.

Scientific Applications:

• Spatial, spectral, and temporal datasets: Analysis of datasets exhibiting spatial, spectral, and temporal dependencies where feature correlations are prevalent.
• Genomic data analysis: Handling correlated high-dimensional features common in genomic studies to improve feature selection and prediction.
• Proteomics: Application to proteomics datasets requiring robust feature-importance estimation amid correlated measurements.
• Interpretability of high-dimensional biological data: Improving feature selection and interpretability for investigations of underlying biological processes in complex datasets.

Methodology:

LAVASET implements an ensemble of decision trees that integrates latent variable analysis derived from feature distance characteristics and local neighborhoods to account for correlations during tree splitting, enabling dimension reduction and adjusted feature-importance estimation.