A selective overview of feature screening for ultrahigh-dimensional data

Area of Science:

• Statistics
• Bioinformatics
• Data Science

Background:

• High-dimensional data are prevalent across scientific fields like genomics, imaging, and finance.
• Analyzing high-dimensional data presents significant statistical challenges, necessitating effective feature and variable selection.
• The sparsity principle, assuming few predictors influence the outcome, underpins many existing variable selection methods.

Purpose of the Study:

• To provide a selective overview of feature screening procedures for ultrahigh-dimensional data.
• To offer insights into constructing marginal utilities for feature screening within specific models.
• To motivate the development and application of model-free feature screening procedures.

Main Methods:

• Review of existing penalized least squares and likelihood methods for sparse model estimation and variable selection.
• Exploration of feature screening strategies tailored for ultrahigh-dimensional data.
• Discussion on the construction of marginal utilities for model-specific feature screening.

Main Results:

• Penalized variable selection methods are effective but face limitations with exponentially growing dimensions.
• Feature screening is essential for managing ultrahigh-dimensional data where dimension p >> sample size n.
• The study highlights the need for both model-specific and model-free screening approaches.

Conclusions:

• Effective feature screening is paramount for the analysis of ultrahigh-dimensional data.
• Understanding marginal utilities aids in developing targeted screening methods.
• Model-free screening procedures offer a flexible alternative for complex, high-dimensional datasets.