Abstract
A natural way of handling imbalanced data is to attempt to equalise the classfrequencies and train the classifier of choice on balanced data. For two-classimbalanced problems, the classification success is typically measured by thegeometric mean (GM) of the true positive and true negative rates. Here we provethat GM can be improved upon by instance selection, and give the theoreticalconditions for such an improvement. We demonstrate that GM is non-monotonicwith respect to the number of retained instances, which discourages systematicinstance selection. We also show that balancing the distribution frequencies isinferior to a direct maximisation of GM. To verify our theoretical findings, wecarried out an experimental study of 12 instance selection methods forimbalanced data, using 66 standard benchmark data sets. The results revealpossible room for new instance selection methods for imbalanced data.