Classification Of Gene Expression Data Based On Extreme Learning Machine

With the human genome project progresses, DNA MicroarrayTechnology is applied to the study of cancer disease. DNA Microarray Technologygenerates a large number of high dimensionality, small sample size of geneexpression data. Therefore, how to find the critical genes that could benefit to cancerclassification or recognition from gene expression data has been very important.This paper studies classification algorithms for gene expression data.Classification that exists extensively in the real-world is a primary task of inductivelearning in machine learning and pattern recognition. In recent years, someclassification techniques have been applied widely in industry, medical, finance,Internet, scientific research and etc. From two aspects of algorithm design andsimulation test, Extreme Learning Machine with fuzzy membership and ExtremeLearning Machine with misclassification cost are investigated experimentally andtheoretically, respectively. In detail, the major contributions of this thesis are asfollowing:1. Fuzzy Extreme Learning Machine. By integrating the fuzzy membership intothe traditional Extreme Learning Machine (ELM), two novel classificationalgorithms, called Fuzzy ELM (FELM) and Binary Fuzzy ELM (BFELM), arerespectively proposed to deal with classification problem when the data samplescontain noise. In FELM, each sample belongs to one class in a certain extent.Whereas, In BFELM, each sample belongs to different classes with differentmemberships. We argue that the generalization performance of BFELM should besuperior to that of FELM by making full use of the information of the data structure.Experiments show that FELM and BFELM achieve better classification results thanthe traditional ELM in the gene expression data cancer classification problem.2. Cost-Extreme Learning Machine. Traditional ELM presumes higher accuracybased on the assumption that all classes have the same cost, and the sample size ofeach class is approximately equal. We propose cost-sensitive versions of ELM to deal with the classification problems with unequal misclassification cost, in whichthe cost matrix of misclassifying each class samples is integrated into the ELMalgorithm or its decision model. The C-ELM algorithms are proposed to furtherimprove the performance of the ELM in terms of total cost.