Evolutionary Instance Resampling for Difficult Data Sets

In the field of machine learning, data set features such as across-class imbalance and class overlap often pose difficulties for classifier algorithms. A number of methods alleviate these difficulties by adjusting the distribution of the data set before classifier construction. Resampling is typically effected by re-weighting, removing, or duplicating instances. Finding a good distribution for the data set, however, is a nontrivial problem. Evolutionary algorithms are frequently used to search for solutions in large, difficult search spaces. In this thesis, four evolutionary approaches are applied to the problem of instance resampling across a variety of data sets and classifier paradigms. In many cases, the evolutionary pre-processing methods are able to produce better classifiers. In particular, an integer-based, one-to-one representation and a cluster-based, real-valued weighting scheme are shown to be beneficial for improving classifier performance on difficult data sets. Index words: genetic algorithms, machine learning, imbalance, undersampling, oversampling, instance selection Evolutionary Instance Resampling for Difficult Data Sets