Learning from imbalanced data refers to the problem of inadequate performance of standard classification algorithms when categorising minority instances, which usually represent rare events of critical importance in the problem domain. Given the complexity and prevalence of this problem, it is not surprising that various approaches to alleviating it are available. The most appropriate are preprocessing procedures, such as feature selection and oversampling, that reduce the problem complexity and the radial basis function network (RBFN) classifier. Although many effective variants of these procedures can be found in the literature, there are certain shortcomings within each. Improvements to these procedures are proposed in the thesis. An extension of bio-inspired wrappers that is based on storing solutions found during search and their subsequent combination according to the impact on classification performance is proposed. Reliance solely on one solution found by the wrapper, which may be overly fitted to the data used for validation, is avoided by storing multiple solutions and combining them in order to identify relevant features for the problem at hand. Also, an improvement of the SMOTE algorithm which determines the neighbourhoods of minority instances and the number of synthetic instances to be created by considering the intrinsic data characteristics is presented. This eliminates the need for parameters that control its operation, rendering it trivial for use compared to the SMOTE algorithm, and especially compared to many of its improved versions. Finally, a new procedure for designing classification models based on RBFN is proposed which aims to find a sequence of networks of favourable classification performance by gradually training networks of increasing complexity using knowledge from previously trained, less complex, networks. To make the most of this knowledge, a node that has beneficial interactions with existing ones is added to the network obtained in the previous step. The proposed improvements were extensively tested on standard datasets that represent various classification problems of different degrees of class imbalance. The analysis of the obtained results led to the conclusion that the proposed improvements enhance the recognition of the minority class and thus improve overall classification performance. Furthermore, it was shown that they outperform common, but also improved variants of these procedures available in the literature.