Interaction screening for high-dimensional heterogeneous data via robust hybrid metrics.

Area of Science:

• Statistics
• Data Science
• Bioinformatics

Background:

• High-dimensional heterogeneous data analysis presents challenges in identifying meaningful interactions.
• Existing methods may struggle with diverse response variable types and capturing complex interaction effects.

Purpose of the Study:

• Introduce a novel model-free interaction screening approach, the hybrid metrics.
• Develop a robust method for analyzing high-dimensional heterogeneous data.
• Enhance interaction selection for various statistical models.

Main Methods:

• The hybrid metrics are based on the variation of conditional joint distribution functions, measuring interaction size and direction.
• The approach is applicable to continuous, discrete, and categorical response variables.
• A fast two-stage procedure is proposed to enforce strong and weak heredity.

Main Results:

• Hybrid metrics effectively screen interactions for classification, response index models, and Poisson regression.
• The method captures both nonlinear category-general and category-specific interaction effects in classification.
• Demonstrated superior performance over existing methods via simulations and a real data example.

Conclusions:

• Hybrid metrics provide a powerful and versatile tool for interaction screening in high-dimensional heterogeneous data.
• The method offers comprehensive insights and precise discovery of interaction information.
• The proposed procedure facilitates efficient and effective implementation.