FunFor

FunFor extends Breiman's random forests to model functional data, predicting curve responses for new observations and selecting significant scalar predictors for analyses of densely and regularly measured functional observations.

Key Features:

• Functional Data Modeling: Models functional responses (densely and regularly measured curves) allowing nonlinear associations and high-order interactions without parametric distributional assumptions.
• Variable Selection and Prediction: Ranks and selects important scalar predictors from large sets while generating curve predictions for new observations.
• Non-Parametric Approach: Operates as a non-parametric method that does not impose restrictive assumptions on data distributions.
• Efficiency and Robustness: Leverages random forest efficiency to detect complex relationships and demonstrates superior performance in variable ranking accuracy, robustness of variable selection, and prediction error reduction in testing.
• Minimal Tuning Requirements: Requires few tuning parameters for model fitting.

Scientific Applications:

• Leaf shape analysis: Applied to biological leaf shape data to model morphological curves and assess predictor importance.
• Genomics: Applicable to genomics datasets where functional responses arise across time or other continuous measurements.
• Environmental science: Suitable for environmental data involving continuous measurements over time or space.
• Time-series and spatially continuous measurements: Supports analyses of time-series or spatially continuous functional observations.

Methodology:

Extends Breiman's random forests to functional-response contexts, predicts curve responses for new observations, ranks and selects scalar predictors, handles nonlinear associations and high-order interactions non-parametrically, is implemented in R, and was evaluated across eight simulation settings and a real-data analysis.