Adaptive Auto-Tuned Distribution-Based Gravitation Sampling for Imbalanced Data Streams Using Bayesian Optimization With Optuna

Imbalanced data streams pose significant challenges due to continuously evolving data distributions, limited minority class samples, and the risk of concept drift. Conventional resampling techniques lack dynamic adaptability and often fail to maintain stable performance in streaming environments. This research proposes an adaptive auto-tuned distribution-based gravitation sampling method integrated with Bayesian optimization using Optuna. Experiments conducted on 21 dataset streams yielded average evaluation metrics ranging from 0.71 to 0.81 with coefficients of variation (CV) below 15% in the low imbalance ratio (IR) category. These results demonstrate that the proposed method can significantly improve classification performance for data streams with low IR and while maintaining high stability across batches.