Imbalance Learning Using Heterogeneous Ensembles

In pattern classification, class-imbalance problem occurs when the number of samples in one of the classes is much larger than those in the others. In such cases, the performance of classifiers is generally poor on the minority class. Ensembles of classifiers are used to tackle this problem where each member is developed using a different balanced dataset. In this approach, one balancing strategy and a classifier prototype is generally used. In order to increase the diversity among the members, bagging and boosting are also considered. In this thesis, the use of heterogeneous ensembles utilizing multiple prototypes and multiple balancing schemes for imbalance learning is addressed. Experiments conducted on 66 datasets have shown that significant improvements can be achieved by employing multiple prototypes. It is also observed that multiple balancing schemes contribute to the performance scores, especially in simple and bagging-based ensembles.