A greedy regression algorithm with coarse weights offers novel advantages.

Area of Science:

• Statistics
• Machine Learning
• Bioinformatics

Background:

• Regularized regression is a standard method for identifying predictors of outcomes.
• Existing methods can be sensitive to data changes and struggle with collinear variables.

Purpose of the Study:

• To present a novel Coarse Approximation Linear Function (CALF) for predictor selection and predictive modeling.
• To develop a method that creates simple, powerful models with robust predictor weighting.

Main Methods:

• CALF is a linear regression strategy applied to normalized data.
• It utilizes non-zero weights of +1 or -1.
• Optimization targets qualitative metrics like p-value, AUC, or Pearson correlation.

Main Results:

• CALF effectively selects important predictors and builds simple, powerful models.
• The method demonstrates stability, with weights largely unaffected by small data changes.
• CALF robustly handles collinear or nearly collinear predictors by selecting at most one.

Conclusions:

• CALF offers a frugal approach to predictor selection and predictive modeling.
• Its use of fixed weights and handling of collinearity enhance model interpretability and stability.
• CALF provides a valuable alternative for risk prediction modeling, especially when qualitative metrics are prioritized.