This project develops a benchmarking framework for wrapper approaches to feature subset selection across datasets and classifiers, including SAGAFS. The goal is to support consistent experiments that compare predictive performance, the number of selected features, and computational cost.
The feature subset selection problem (FSSP) asks which subset of available features best supports a classifier’s predictions. A feature’s usefulness depends on the other selected features and the learning algorithm; the best subset need not be unique, and its size is not known in advance. Because the number of possible subsets grows exponentially, exhaustive search quickly becomes impractical.
Wrapper approaches evaluate candidate subsets by training and assessing a classifier. Their three components are a search strategy, a performance estimation method, and a learning algorithm. This framework provides a common setting for comparing those choices.
For peptide classification, molecular descriptors quantify properties such as amino acid composition, charge, hydrophobicity, and structural tendencies. Selecting a useful combination helps build compact representations for distinguishing antimicrobial from non-antimicrobial peptides. SAGAFS (Species Adaptive Genetic Algorithm for Feature Selection) searches subsets using variable-length representations and evaluates predictive performance while favoring smaller subsets when performance is tied.
Bohdan Hrotovytskyy
Master’s student · California State University, Los Angeles
Beltran, J. A., Del Rio, G., & Brizuela, C. A. (2020). An automatic representation of peptides for effective antimicrobial activity classification. Computational and Structural Biotechnology Journal, 18, 455–463.