Improved Oversampling Algorithm for Imbalanced Data Based on K-Nearest Neighbor and Interpolation Process Optimization

Abstract

The problems of imbalanced datasets are generally considered asymmetric issues. In asymmetric problems, artificial intelligence models may exhibit different biases or preferences when dealing with different classes. In the process of addressing class imbalance learning problems, the classification model will pay too much attention to the majority class samples and cannot guarantee the classification performance of the minority class samples, which might be more valuable. By synthesizing the minority class samples and changing the data distribution, unbalanced datasets can be optimized. Traditional oversampling algorithms have problems of blindness and boundary ambiguity when synthesizing new samples. A modified reclassification algorithm based on Gaussian distribution is put forward. First, the minority class samples are reclassified by the KNN algorithm. Then, different synthesis strategies are selected according to the combination of the minority class samples, and the Gaussian distribution is used to replace the uniform random distribution for interpolation operation under certain classification conditions to reduce the possibility of generating noise samples. The experimental results indicate that the proposed oversampling algorithm can achieve a performance improvement of 2∼8% in evaluation metrics, including G-mean, F-measure, and AUC, compared to traditional oversampling algorithms.