Entropy-based Criteria Dealing with the Ties Problem in Gene Selection

Gene selection is one of the major issues in microarray gene expression data analysis. Among existing gene selection techniques, wrapper methods usually produce better selection results than filter methods and embedded methods. The wrapper methods employ a predefined classification algorithm and evaluate the goodness of gene subsets in terms of classification error, which is usually obtained through error counting. Due to the high dimensionality and small sample size of gene expression data, counting-based evaluation criteria could result in severe ties problem which induces selection uncertainty and poor robustness. In this study, we will demonstrate the existence of the ties problem by well designed experiments. In addition, two continuous evaluation criteria based on Renyi's entropy and support vector machines (SVMs) were proposed. Experiment results show that the proposed SVM-Renyi's entropy criteria could well overcome the ties problem and provide gene subsets leading to improved classification accuracy.