Oversampling Algorithm Based on Spatial Distribution of Data Sets for Imbalance Learning

Abstract

Imbalance problem is widespread in machine learning. Most learning algorithms can’t get satisfied performance when they are applied on imbalance data sets, because they can be deteriorated by this problem easily. This paper proposed SDSMOTE method which captures the spatial distribution of imbalance data sets, and changes the tendency of learning algorithm by over sampling by oversampling according to the recognition difficulty. Experiments on 5 UCI data sets validate the effectiveness of this oversampling algorithm.