J.M.C. Sousa - Feature Selection using Ant Colony Optimization: Applications in Health Care
Feature selection is a scientific research area that is emerging particularly fast. Although there are more or less complex methods to perform feature selection, there is no established method that guarantees an optimal or simply a good solution. In this seminar, two different goals are defined for the objective function: maximize model accuracy and minimize the number of used features. Metaheuristics can perform a global search of the feature search space, and as so are very appealing. The metaheuristic ant colony optimization is applied to the feature selection problem. The concept of exchanging information using two different ant colonies is introduced. Each colony is responsible for one of the goals formulated in the feature selection optimization problem. However, the goals are contradictory and have different significance throughout the search space. To cope with this problem, fuzzy criteria are defined for each goal and the effectiveness of the approach is shown. The proposed approaches are applied to several benchmark data sets. The main application is in health care; namely, to predict the outcomes of septic shock patients. The goal is to estimate, as accurately as possible, the outcome (survived or deceased) of these patients. Results show that the presented approaches outperform any previous solutions, specifically in terms of sensitivity.