In tackling the issues of overfitting and limited generalization in machine learning models, especially with imbalanced and limited data, oversampling methods have been widely used to enhance model discrimination by generating synthetic minority samples. However, these methods often struggle to capture the expressive nuances of the minority class due to noisy interference. In response to this challenge, we propose a three-way oversampling (3WOS) method that partitions data into three regions via density clustering, then synthesizes samples in purified regions to avoid noise. The 3WOS operates in two stages: first, a three-way density clustering (3WDC) process uses density ratios to divide data and identify the FringeRegion (Fr) as a source of benchmark representative samples, aiding in similar data recognition and noise mitigation. Second, a data synthesis strategy creates a purified region from the representative samples, where the FringeRegion provides meaningful, discriminative data for synthesis through linear interpolation. This approach avoids data redundancy and overlap, ensuring a more robust minority representation. Extensive experiments validate 3WOS effectiveness, demonstrating significant improvements in model performance, particularly in enhancing minority class representation.