In supervised learning, rare events are difficult to identify due to their infrequency in the training data. This is a huge issue since misclassifying rare events can be very costly. For example, failing to detect a serious financial fraudulent transaction can cause huge losses to an organization. This type of situation can be referred to as an imbalanced data classification problem, where one of the classes represents a very small minority of the data and the most prevalent class is called the majority class. To tackle the imbalance issue, many approaches have been proposed over the years to cope with imbalanced data classification, most of which have been based on data re-sampling, algorithm adaptions, cost-sensitive learning, ensemble learning, as well as different combinations of these approaches.
Despite the attraction that imbalanced data is receiving in literature, there are still many issues relevant to this topic that are not yet addressed. Uncertainty is one of those issues, which is a common problem in supervised learning. In fact, it can be found in every stage of the learning process, from data preprocessing to model selection. In imbalanced classification, we can distinguish different sources of uncertainties. To deal with them, this thesis will focus on the development of belief function theory based methods to tackle challenging class-imbalanced problems.