Balanced accuracy for feature subset selection with genetic algorithms

摘要: The relevance of a set measured features describing labeled patterns within problem domain affects classifier performance. Feature subset selection algorithms employing wrapper approach typically assess the fitness feature simply as accuracy given over available using candidate set. For datasets with many for some classes and few others, relatively high may be achieved by labeling unknown according to largest class. wrappers that only emphasize follow this bias. Class bias mitigated emphasizing well-balanced during optimization algorithm. This paper proposes adding selective pressure balanced mitigate class evolution. Experiments compare performance genetic various functions varying in terms accuracy, balance, parsimony. Several including greedy, genetic, filter, hybrid filter/GA approaches are then compared best function. experiments employ naive Bayes public datasets. results suggest improvements balance size can made without compromising overall or run-time efficiency.