Integration of rank-partition and sequential wrapper techniques for feature selection of breast cancer microarray data

• Main Author: Abdulwahhab, Ahmed Abbas
• Format: Thesis
• Language: English
• Published: 2015
• Online Access: http://psasir.upm.edu.my/id/eprint/56209/1/FK%202015%203RR.pdf; http://psasir.upm.edu.my/id/eprint/56209/

A deoxyribonucleic acid (DNA) microarray has the ability to record huge amount of genetic information simultaneously. Previous researches have shown that this technology can be helpful in the classification of cancers and their treatments outcomes. This has encouraged information technology engineers to cooperate in microarray data analysis for enhancing medicine and biology technologies. Typically, cancer-related microarray data are consisted of high dimensional gene expression levels (as features) for a limited number of samples. This characteristic in the structure of microarray data causes the phenomenon known as the curse of dimensionality, which is a particularly problem for standard classification models. It contradicts to the required ratio of samples to genes which should be much greater than 1 and it makes the direct application of machine learning techniques inefficient. Consequently, gene selection techniques have become a crucial element in the classification of microarray data. Based on previous researches in the context of microarray data classification, the results obtained from the classification of breast cancer data have the lowest accuracy among them. Therefore, this study was aimed in improving the classification accuracy of clinical outcomes for breast cancer by gene expression profiling. Filter and wrapper for gene selection are the main techniques in many existing microarray data analysis. Promising results obtained from filter-wrapper techniques have led to the design of a proposed model for this study. A gene selection model that integrates rank-partition and sequential wrapper was designed to find optimal subset of the most informative genes that enhances the predictive power of gene expression profiling. Evaluation of the obtained results for breast cancer data set demonstrates that the proposed integrated model achieved the objective in finding the optimal subset of the most informative genes that has the predictive power of 87% accuracy compared to 83% of the original study and 77% of the shrunken centroid method.