Abstract
<p>Accurate variable selection is central to building predictive models for two-group classification. Genetic algorithms (GAs) are a flexible family of metaheuristic search methods for this task, and a defining feature of that flexibility is that the same evolutionary search can be paired with many different internal classification models and fitness functions. This flexibility, however, introduces researcher-driven heterogeneity whose consequences for variable selection performance are largely unexamined. This study has two parts. First, a systematic review, conducted following PRISMA 2020 guidance, documents the internal models and fitness functions used in published GA-based variable selection for classification, and quantifies the heterogeneity in these choices. Second, a Monte Carlo simulation holds the GA search strategy fixed while fully crossing the three most common classifiers identified in the review (support vector machine, k-nearest neighbors, and logistic regression) with the three most common fitness function types (classification accuracy, a multi-objective accuracy criterion penalized by subset size, and area under the ROC curve). The nine resulting implementations are evaluated across 108 data-generating conditions that vary sample size, number of candidate predictors, proportion of true predictors, inter-predictor collinearity, and class balance, with 1,000 replications per condition. This preprint includes pilot simulations of a subset of these conditions. Performance is summarized along four dimensions: out-of-sample classification performance (estimated on an independent test partition), variable selection recovery relative to the known true predictor set, selected subset size, and computation time. Classification metrics and recovery metrics that are both aligned and unaligned with each fitness function are reported, so that trade-offs between optimization criteria can be examined directly.</p>