Feature selection for multi-purpose predictive models: A many-objective task

Abstract

The target of machine learning is a predictive model that performs well on unseen data. Often, such a model has multiple intended uses, related to different points in the tradeoff between (e.g.) sensitivity and specificity. Moreover, when feature selection is required, different feature subsets will suit different target performance characteristics. Given a feature selection task with such multiple distinct requirements, one is in fact faced with a very-many-objective optimization task, whose target is a Pareto surface of feature subsets, each specialized for (e.g.) a different sensitivity/specificity tradeoff profile. We argue that this view has many advantages. We motivate, develop and test such an approach. We show that it can be achieved successfully using a dominance-based multiobjective algorithm, despite an arbitrarily large number of objectives. © 2010 Springer-Verlag.