Multi-Objective Genetic Programming for Classification with Unbalanced Data

Existing learning and search algorithms can suffer a learning bias when dealing with unbalanced data sets. This paper proposes a Multi-Objective Genetic Programming (MOGP) approach to evolve a Pareto front of classifiers along the optimal trade-off surface representing minority and majority class accuracy for binary class imbalance problems. A major advantage of the MOGP approach is that by explicitly incorporating the learning bias into the search algorithm, a good set of well-performing classifiers can be evolved in a single experiment while canonical (single-solution) Genetic Programming (GP) requires some objective preference be a priori built into a fitness function. Our results show that a diverse set of solutions was found along the Pareto front which performed as well or better than canonical GP on four class imbalance problems.