Adaptive subspace sampling for class imbalance processing
Yuting Liu, Nikhil Ranjan Pal, Shang-Lin Wu, Tsung-Yu Hsieh, Chin‐Teng Lin · 2016
This paper presents a novel oversampling technique that addresses highly imbalanced data distribution. At present, the imbalanced data that have anomalous class distribution and underrepresented data are difficult to deal with through a variety of conventional machine learning technologies. In order to balance class distributions, an adaptive subspace self-organizing map (ASSOM) that combines the local mapping scheme and globally competitive rule is proposed to artificially generate synthetic samples focusing on minority class samples. The ASSOM is conformed with feature-invariant characteristics, including translation, scaling and rotation, and it retains the independence of basis vectors in each module. Specifically, basis vectors generated via each ASSOM module can avoid generating repeated representative features that offer nothing but heavy computational load. Several experimental results demonstrate that the proposed ASSOM method with supervised learning manner is superior to other existing oversampling techniques.