A comparative study on sampling techniques for handling class imbalance in streaming data
Hien M. Nguyen, Eric Wallace Cooper, Katsuari Kamei · 2012
Sampling is the most popular approach for handling the class imbalance problem in training data. A number of studies have recently adapted sampling techniques for dynamic learning settings in which the training set is not fixed, but gradually grows over time. This paper presents an empirical study to compare over-sampling and under-sampling techniques in the context of data streaming. Experimental results show that under-sampling performs better than over-sampling at smaller training set sizes. All sampling techniques, however, are comparable when the training set becomes larger. This study also suggests that a multiple random under-sampling (MRUS) technique should be a good choice for applications with imbalanced and streaming data, because MRUS is the most effective while still keeping a high speed.