A new approach to classification with the least number of features

The ever increasing complexity of real-world machine learning tasks requires more and more sophisticated methods to deal with datasets that contain only very few relevant features but many irrelevant noise dimensions. In practise, these scenarios often arise in the analysis of biological datasets, such as tissue classification using microarrays [1], identification of disease-specific genome mutations or distinction between mental states using functional magnetic resonance imaging [2]. It is well-known that a large number of irrelevant features may distract state-of-the-art methods, such as the support vector machine. Thus, feature selection is often a fundamental preprocessing step to achieve proper classification results, to improve runtime, and to make the training results more interpretable.