Abstract
In the recent years, many applications in machine learning involve an increasingly large number of features and samples, which poses new challenges to many learning algorithms. To address these challenges, ensemble learning methods, which uses multiple base learners, have been proposed to achieve better predictive performance. This thesis covers a range of topics in ensemble classification, including multi-objective and semi-supervised heterogeneous classier ensembles. We first present an empirical study on heterogeneous classifier ensembles, which confirms that heterogeneous ensembles outperform homogeneous ones and single classifiers. Secondly, we present a multi-objective ensemble generation method, which creates a group of members so that the diversity among the base learners could be explicitly maintained. The third topic of this thesis is a feature selection method for data that has a large number of features. By using the modified competitive swarm optimizer as the search algorithm, we are able to considerably reduce the number of features and at the same time improve the classifiers' generalisation performance. Finally, we present a novel semi-supervised ensemble learning algorithm, termed Multi-Train, that uses semi-supervised learning algorithms to learn from unlabelled data.