Mining Gene Expression Data

Micro-array data has enabled us to obtain an overview of the cell by measuring the expression levels of thousands of genes simultaneously. It has also opened the possibility for predicting possible cancerous cells accurately. Since these data sets are very large, we need to utilize machine learning techniques to analyse them efficiently. However, micro-array data also poses challenging problems for machine learning algorithms. Previously, analysis on micro-array data was inferred simply from the micro-array data itself and ignored all the information surrounding it. Neglecting information is however contrary to standard machine learning wisdom. Therefore, the main goal of this work was to investigate the hypothesis that prediction could be improved by utilizing additional information. To enable linking of the micro-array data with its additional information in WEKA, a new space-efficient linked instance representation had to be invented. This new representation enabled effective performance comparisons of various machine learning algorithms on three different sets of gene expression data with a varying amount of additional linked-in information. Generally, we found SMO to be the single best classifier and voting of the top three classifiers usually improved performance over the single best classifier. Globally, best results were usually achieved using the linked-in additional information, thus confirming our original hypothesis.