Data Preprocessing Strategies in Imbalanced Data Classification

Learning from imbalanced data has been a research topic studied for many years. There are two main approaches used today - data-level and algorithm- level methods. We set out to study resampling methods which belong to the category of data-level methods. These methods modify the training part of a dataset as opposed to algorithm-level methods, which modify a classifier itself. Resampling methods are further divided into oversampling and un- dersampling methods. It is challenging to know which group of methods performs better and which algorithms stand out the most. We conducted an experiment of unseen scale. We systematically and robustly compared sixteen preprocessing methods over eighteen imbalanced datasets and sum- marised the results in this thesis. The results show that oversampling meth- ods outperformed most undersampling methods in both performance and preprocessing time.