A Smooth Oversampling Technique Using SLP

The class imbalance problem occurs when the number of samples in a given dataset is not balanced according to the class labels. In a binary classification problem of two classes, one class contains a minority of samples while the other has a majority. The commonly utilized SMOTE technique generates synthetic samples in the middle of a pair of original samples in the feature space to oversample the minority class. The midlines and randomness characteristics of this technique and its extensions capture the local characteristics of the minority class. Besides, these techniques do not significantly improve the classification problem’s performance. Accordingly, this paper proposes a new technique for oversampling using a single-layer perceptron that captures the general characteristics of the data with low resource consumption. The proposed technique uses the weights trained for the network as input to other networks to generate new samples within the min-max range. The results showed that the proposed technique overperformed the SMOTE techniques, as reported using several machine learning techniques, including decision tree, random forest, support vector machine, and K-nearest neighbors. The experiments are conducted using the highly imbalanced credit card transaction dataset, which contains 492 fraudulent transactions out of 284,807 total transactions. The random forest reported the best results, in which the proposed technique improves the recall and the f-measure while maintaining full precision, similar to SMOTE.