A Hybrid Model for Imbalanced Data Classification Using Dynamic Threshold Tuning and Particle Swarm Optimization‐Enhanced Kernel Transformations

ABSTRACT In real‐world applications, acquiring the necessary labeled data for training machine learning models can be expensive, requiring significant human effort and time. This study proposes a learning framework that utilizes both labeled and unlabeled data to overcome the challenges of imbalanced data classification. A feature ranking method is introduced in the preprocessing stage to remove redundant and irrelevant features. The model employs a dynamic confidence‐based filtering mechanism to selectively incorporate high‐confidence unlabeled instances, thereby enhancing the learning process without using noisy or uncertain data and reduces the dependency on fully labeled datasets. Our framework includes dynamic threshold tuning, kernel‐based transformation using both linear and radial basis function kernels, and fine‐tuning the model through particle swarm optimization for optimal performance. Experimental results on datasets from the knowledge extraction based on evolutionary learning (KEEL) repository demonstrate that the proposed model consistently outperforms support vector machines, particularly in imbalanced datasets irrespective of the imbalance ratio with an average improvement of 8%–10%. The suggested model significantly improves classification performance compared to the supervised learning approach, making it a viable solution for real‐world imbalanced data classification problems. Moreover, the analysis of the threshold tuning reveals that as the adaptive threshold approaches an optimal range around 0.5, performance of the model improves. As a future work, we can extend this framework to multi‐class classification problems, by exploring the use of multi‐objective kernel functions, and thereby improving scalability for large datasets.