Mathematical approachesto synthetic data sampling methods

In this thesis, a method is proposed to improve performance losses in machine learning algorithms of unevenly distributed datasets. Many studies have been done to reduce or completely remove the imbalance in datasets (RUS, ROS, SMOTE). Similarly, in the developed method, existing samples belonging to the minority class were reproduced synthetically and the datasets were balanced. For the resampling process, the nearest neighbors for all data points were determined using the Euclidean distance metric among the samples belonging to the minority class. Different from the other methods, a desired number of new synthetic samples were created using the Weighted Geometric Average, among these neighbors, in a sufficient number among the nearest neighbors. Significant improvements were observed in the machine learning performance of datasets balanced in this way, compared to the methods compared.