Enhanced Feature Selection Techniques for Classification of Microarray Gene Expression Data

Abstract

Thanks to the breakthroughs that have taken place in bioinformatics during recent years. Data in the field of bioinformatics is accumulating at an unseen pace. Data-intensive, large scale biological problems from the computational point of view were addressed in bioinformatics. These data are still analyzed using some old statistical methods. Microarray technology has been used extensively to study the gene expression of Deoxy Ribonucleic Acid (DNA) data for the diagnosis and prognosis of diseases, including cancer in the organism. However, advancement in microarray technology has facilitated the researcher to the next level for analyzing the gene expression datasets. Hence, microarray technology has evolved as one of the powerful tools for analysing gene expression levels for an organism, that facilitated the researcher to apply Machine learning algorithms like classification for analyzing and extracting exciting patterns of gene expression. However, microarray datasets contain a very massive amount of features( genes) but comparatively small samples size, because of the mentioned characteristics of microarray data, the efficiency of machine learning algorithms are taxed. Hence, there is a need for feature engineering for achieving better outcomes of machine learning algorithms for obtaining interesting gene expression patterns. This thesis focuses on improvising the efficiency of the machine learning algorithm by using feature engineering on microarray datasets. The works presented in this thesis are improvised feature engineering techniques with the help of proper resampling with efficient feature engineering and hybridization of mRMR and AGA method. The experiments were conducted using RFEVSS and RSRV as an efficient feature selection on four microarray datasets, which was the first contribution. The microarray gene expression data common characteristics are high dimension along with a small newlinesample size. The performance of the learning algorithms is never up to mark because of these characteristics. Another c