<p>Oversampling is a common approach to mitigating class imbalance by adjusting data distribution, facilitating standard classifier training. However, conventional oversampling methods struggle with noise and intra-class imbalance, leading to suboptimal synthetic sample generation. To tackle these issues, this paper proposes an improved Synthetic Minority Over-sampling Technique (SMOTE) based on three-way decision theory and k-means++ clustering, termed the Three-way k-means Oversampling (TWKO) algorithm. TWKO applies k-means++ to partition the entire dataset, and further subdivides each cluster into three domains—positive, negative, and boundary—based on the imbalance ratio (IR). To better control the strictness of these domain thresholds, a balance exponent n is introduced, which sharpens or smooths the decision boundaries by adjusting the influence of IR during three-way domain partitioning. After noise removal, distinct sampling strategies are employed within each domain to generate synthetic samples more effectively. TWKO was validated on nine UCI and four combined datasets using six classifiers, benchmarked against five state-of-the-art algorithms. Experimental results confirm that TWKO consistently outperforms five benchmark methods, demonstrating its robustness in addressing class imbalance.</p>