A framework for feature selection in high-dimensional domains

The introduction of DNA microarray technology has lead to enormous impact in cancer research, allowing researchers to analyze expression of thousands of genes in concert and relate gene expression patterns to clinical phenotypes. At the same time, machine learning methods have become one of the dominant approaches in an effort to identify cancer gene signatures, which could increase the accuracy of cancer diagnosis and prognosis. The central challenges is to identify the group of features (i.e. the biomarker) which take part in the same biological process or are regulated by the same mechanism, while minimizing the biomarker size, as it is known that few gene expression signatures are most accurate for phenotype discrimination. To account for these competing concerns, previous studies have proposed different methods for selecting a single subset of features that can be used as an accurate biomarker, capable of differentiating cancer from normal tissues, predicting outcome, detecting recurrence, and monitoring response to cancer treatment. The aim of this thesis is to propose a novel approach that pursues the concept of finding many potential predictive biomarkers. It is motivated from the biological assumption that, given the large numbers of different relationships which are possible between genes, it is highly possible to combine genes in many ways to produce signatures with similar predictive power. An intriguing advantage of our approach is that it increases the statistical power to capture more reliable and consistent biomarkers while a single predictor may not necessarily provide important clues as to biological differences of interest. Specifically, this thesis presents a framework for feature selection that is based upon a genetic algorithm, a well known approach recently proposed for feature selection. To mitigate the high computationally cost usually required by this algorithm, the framework structures the feature selection process into a multi-step approach which combines different categories of data mining methods. Starting from a ranking process performed at the first step, the following steps detail a wrapper approach where a genetic algorithm is coupled with a classifier to explore different feature subspaces looking for optimal biomarkers. The thesis presents in detail the framework and its validation on popular datasets which are usually considered as benchmark by the research community. The competitive classification power of the framework has been carefully evaluated and empirically confirms the benefits of its adoption. As well, experimental results obtained by the proposed framework are comparable to those obtained by analogous literature proposals. Finally, the thesis contributes with additional experiments which confirm the framework applicability to the categorization of the subject matter of documents.