Abstract A Feature Selection Method for Gene Expression Data with Thousands of Features

We examine feature selection algorithms for handling data sets with many features. We introduce the Random Forest approach and the Scatter Search method as algorithms for selecting relevant features. We then present a general 3-step approach for data sets with thousands of features. Step 1 consists of a Feature Space Reduction process where the majority of irrelevant features are eliminated. This is followed by a Feature Subset Optimization process (Step 2) where we attempt to search for a nearoptimal and small subset in the reduced space. These subsets are then evaluated in Step 3 for their efficiency in classifying the samples in the data. We tested our approach with two gene expression data sets. One of the data sets has 62 samples and 2000 features, and the other has 72 samples and 7129 features. We achieved results in support of our methods. The feature selection algorithms and machine learning techniques that we studied have important applications in the medical community. The ability to classify cell and cancer types can lead to significant changes in the diagnosis and treatment of many diseases.