Handling Class Imbalance in High-Dimensional Biomedical Datasets

When dealing with biomedical data, the first and most challenging issue is often the huge dimensionality, i.e. the presence of a very high number of features for each of the problem instances at hand. A vast literature is available on different dimensionality reduction techniques that can be suitable for handling such kind of data, with a special focus on feature selection algorithms that allow to discard uninformative/useless features. In most cases, however, the dimensionality issue is addressed without a joint consideration of other potential problems in the data, including an imbalanced class distribution that may hinder the construction of effective classification models. Class imbalance, in turn, has been mostly treated in literature as an independent problem, especially in application fields where the number of features is not so critical. But several biomedical datasets are both high-dimensional and class-imbalanced, so there is a strong need for designing and evaluating learning strategies that can properly deal with both the issues simultaneously. In this work, we experiment with using feature selection techniques in conjunction with sampling-based class balancing methods and cost-sensitive classification, in order to gain insight into the most effective strategies to use when dealing with such complex data.