Cost-Constrained feature selection in binary classification: adaptations for greedy forward selection and genetic...

Area of Science:

• Biotechnology
• Statistical analysis
• Bioinformatics

Background:

• Biotechnology advances present complex statistical challenges in biomarker discovery from high-dimensional data.
• Feature selection is vital for managing numerous biomarker candidates in biomedical data analysis.
• Biomarker candidate costs, financial or otherwise, are an important but under-researched consideration in medical applications.

Purpose of the Study:

• To extend existing feature selection methods (greedy forward selection, genetic algorithms) to incorporate cost constraints.
• To evaluate the performance of these adapted methods against baseline alternatives under budget limitations.

Main Methods:

• Development of modified greedy forward selection and genetic algorithms to control feature costs.
• Simulation studies involving binary classification tasks to compare methods.
• Assessment of predictive performance, run-time, and relevant feature detection rates.

Main Results:

• Proposed cost-aware feature selection methods outperformed baseline alternatives under predefined budget constraints.
• A slight performance drop was observed in the adapted greedy forward selection without a budget constraint, which can be mitigated by a hyperparameter adjustment.
• Minor differences in performance were noted between the proposed methods themselves.

Conclusions:

• Standard feature selection algorithms are often inadequate for identifying optimal biomarker subsets when cost budgets are imposed.
• Adaptations to feature selection algorithms, as proposed, are effective in addressing feature cost scenarios and budget constraints in biomarker selection.