Global Data Distribution Weighted Synthetic Oversampling Technique for Imbalanced Learning

Imbalanced learning is a common problem in data mining. There different distribution of samples among other classes the imbalanced datasets. It's challenge for standard algorithms designed balanced class distributions. Although there are various strategies to solve this problem, generating artificial achieve relatively universal rather than directly modifying specific classification algorithms. The oversampled can be combined with any user-specified algorithm without restrictions. In paper, we present novel oversampling method, Global Data Distribution Weighted Synthetic Oversampling Technique (GDDSYN). By applying clustering, optimizing selection criteria minority that used generate synthetic samples, avoiding more noise samples. GDDSYN assigns weights number tackle within-class imbalance and between-class simultaneously, according informative level sample sparsity cluster which belongs. use scores Silhouette Coefficient Mutual Information helps k-means set reasonable clusters majority respectively so clustering effect guaranteed. Next, by using information, samples' generation path improved avoid overlap. Additionally, has been evaluated extensively on 10 real-world sets. empirical results show our method outperforms or comparable some existing methods terms assessment metrics when generated used.