Review of the effect of feature selection for microarray data on the classification accuracy for cancer data sets

摘要

DNA microarrays can be used to monitor the expression level of thousands of genes simultaneously and gene microarray data can be used in cancer diagnosis and classification. Many machine learning techniques have been developed for computational analyses of microarray data. A common difficulty for all techniques is the large number of genes compared to the small sample size which has a negative impact on their speed and accuracy. To overcome these limitations, feature selection techniques are applied to distinguish between significant and redundant or irrelevant genes. Feature selection methods are used for two main goals. The first is to identify the relationship between specific diseases and genes. The second is to examine a compact set of discriminative genes to develop a pattern classifier with good generalizability and limited complexity. Here, we review different feature selection methods for cancer microarray data sets and analyze their accuracy. We describe methods commonly used for selecting significant features including filters, wrappers and embedded methods, categorized according to their experimental methodology. We then compare the classification accuracy of the methods for various cancer data sets and their time complexity to make some suggestions regarding the use of suitable methods for cancer data sets.