Scatter search for high-dimensional feature selection using feature grouping

In feature selection tasks, finding the optimal subset of features is unfeasible due to the increase of the search space with the dimensionality. In order to reduce the complexity of the space, feature grouping approach aims to generate subsets of correlated features. In this context, evolutionary algorithms have proven to achieve competitive solutions. In this work we propose a novel Scatter Search (SS) strategy that uses feature grouping to generate a population of diverse and high quality solutions. Solutions are evolved by applying random mechanisms in combination with the feature group structure to maintain the diversity and the quality of the solutions during the search. We test the proposed strategy on high dimensional data from biomedical domains and compare the performance against the first adaptation of the SS to the feature selection problem. Results show that our proposal is able to find smaller subsets of features while keeping a similar predictive power of the classifier models. Finally, a case of study regarding melanoma skin cancer is analysed using the proposed strategy.