A Constructive Method for Data Reduction and Imbalanced Sampling

A large number of training data lead to high computational cost in instanced-based classification. Currently, one of the mainstream methods to reduce data size is to select a representative subset of samples based on spatial partitioning. However, how to select a representative subset while maintaining the overall potential distribution structure of the dataset remains a challenge. Therefore, this paper proposes a constructive data reduction method called Constructive Covering Sampling (CCS) for classification problems. The CCS does not rely on any relevant parameters. It iteratively partitions the original data space into a group of data subspaces, which contains several samples of the same class, and then it selects representative samples from the data subspaces. This not only maintains the original data distribution structure and reduces data size but also reduces problem complexity and improves the learning efficiency of the classifier. Furthermore, CCS can also be extended as an effective undersampling method (CCUS) to address class imbalance issues. Experiments on 18 KEEL and UCI datasets demonstrate that the proposed method outperforms other sampling methods in terms of F-measure, G-mean, AUC and Accuracy.