Guiding ensemble methods by using genesets

Maja Žbogar (2014) Guiding ensemble methods by using genesets. EngD thesis.

Preview

PDF Download (1235Kb)

Abstract

The following bachelor thesis is mainly focused on the field of ensemble machine learning. In contrast to usual machine learning methods, where one strives to find a single best performing model, ensembles are based on the idea of combining multiple base models. They are generally considered to be more successful, than any of their constituent parts. The theoretical part of the thesis explores possible reasons for this superior performance and presents some of the most influential frameworks on ways of building ensembles. Core focus of the empirical research is building its foundations on the philosophy of selecting subsets of input variable space, proposed by the random subspace theory framework, but instead of using randomly selected variables, it explores the idea of using subsets of features, sharing some meaningful connection. Meaningful sets of features, required in order to build the base models, are obtained by using predefined gene sets. Modified ensemble models are built and tested on several different DNA microarray data sets, analysed by using GSEA analysis. The performance of the suggested modifications is compared to the results achieved by using other learning methods. Results indicate, that suggested approach does not yield the desired performance improvements. Possible reasons for the absence of results are investigated and some main findings of the conducted research are highlighted.

Actions (login required)