An efficient model-free approach to interaction screening for high dimensional data.

Area of Science:

• Statistics
• Data Science
• Machine Learning

Background:

• High-dimensional data analysis presents challenges in identifying complex interactions.
• Existing methods may struggle with diverse data types or weak main effects.

Purpose of the Study:

• To introduce a novel, model-free interaction screening procedure (MCVIS) for high-dimensional data.
• To develop a method capable of detecting interactions robustly, even with weak parental main effects.

Main Methods:

• The MCVIS procedure utilizes a proposed MCV index to quantify interaction effect importance.
• The method is fully nonparametric, accommodating discrete, categorical, and continuous covariates.
• It handles both categorical and continuous responses, including missing data, and is robust to heavy-tailed distributions.

Main Results:

• MCVIS demonstrates sure screening and ranking consistency properties, enabling effective dimension reduction.
• The procedure is computationally feasible, simple, and fast to implement.
• Extensive simulations and real-world data analyses confirm its effectiveness and broad applicability.

Conclusions:

• The MCVIS procedure offers a powerful and versatile tool for interaction screening in high-dimensional settings.
• Its nonparametric nature and robustness make it suitable for complex and heterogeneous datasets.
• MCVIS facilitates efficient dimension reduction while preserving crucial interaction information.