Ensemble learning and pruning in multi-objective genetic programming for classification with unbalanced data

Abstract

Machine learning algorithms can suffer a performance bias when data sets are unbalanced. This paper develops a multi-objective genetic programming approach to evolving accurate and diverse ensembles of non-dominated solutions where members vote on class membership. We explore why the ensembles can also be vulnerable to the learning bias using a range of unbalanced data sets. Based on the notion that smaller ensembles can be better than larger ensembles, we develop a new evolutionary-based pruning method to find groups of highly-cooperative individuals that can improve accuracy on the important minority class. © 2011 Springer-Verlag.