Functional random forests for curve response.

Area of Science:

• Statistics
• Machine Learning
• Data Science

Background:

• Increasing demand for advanced statistical methods to handle complex functional data.
• Limitations of traditional methods in capturing nonlinear associations in functional data.

Purpose of the Study:

• To introduce a novel functional random forests (FunFor) approach for modeling densely and regularly measured functional data responses.
• To extend Breiman's traditional random forests to accommodate functional data.
• To develop a method capable of predicting curve responses and selecting important variables from scalar predictors.

Main Methods:

• Development of the functional random forests (FunFor) approach, extending traditional random forests.
• Non-parametric modeling without distributional assumptions.
• Implementation in an R package named 'FunFor'.

Main Results:

• FunFor demonstrates excellent performance in various simulation settings and a real-data analysis.
• Successfully ranks true predictors as most important, achieving robust variable selection.
• Outperforms three other relevant approaches in prediction errors and variable selection.

Conclusions:

• The FunFor approach effectively models functional data, capturing complex relationships including nonlinearities and interactions.
• Its non-parametric, distribution-free, and model-free nature makes it widely applicable across fields.
• The 'FunFor' R package facilitates its use in biological and other data analyses.