Discovering Minority Sub-clusters and Local Difficulty Factors from Imbalance Data

Learning classifiers from imbalanced data is particularly challenging when class imbalance is accompanied by local data difficulty factors, such as outliers, rare cases, class overlapping, or minority class decomposition. Although these issues have been highlighted in previous research, there have been no proposals of algorithms that simultaneously detect all the aforementioned difficulties in a dataset. In this paper, we put forward two extensions to popular clustering algorithms, ImKmeans and ImScan, and one novel algorithm, ImGrid, that attempt to detect minority sub-clusters, outliers, rare cases, and class overlapping. Experiments with artificial datasets show that ImGrid, which uses a Bayesian test to join similar neighboring regions, is able to re-discover simulated clusters and types of minority examples on par with competing methods, while being the least sensitive to parameter tuning.