<p>Data stream learning is typically affected by class distribution imbalance, which causes learning algorithms to be severely biased toward the majority class due to the scarcity of minority samples, thereby significantly reducing their ability to identify minority instances. At the same time, data streams commonly exhibit concept drift, defined as the phenomenon where the input–output relationship learned by the model changes over time, requiring models to maintain ongoing adaptability. The challenges of multi-class imbalance and concept drift greatly weaken the generalization ability and classification accuracy of models. To address these issues, an ensemble classification algorithm tailored for multi-class imbalanced data streams is proposed. First, an expert group validation strategy based on feature-importance updates is designed to effectively capture the dynamic characteristics of multi-class imbalanced data streams under concept drift, thereby enhancing the adaptability of the ensemble model. Second, an adaptive oversampling strategy is introduced, which flexibly adjusts the sample generation process to produce instances consistent with the current data distribution, thus improving the recognition ability of the model for minority classes in drifting environments. Finally, a sample weighting strategy based on Focal Loss and boundary proximity is developed to dynamically adjust the training weights of samples, thereby strengthening the model’s focus on minority or highly uncertain instances. Experimental results demonstrate that the proposed algorithm achieves superior performance on various multi-class imbalanced data streams with concept drift, outperforming existing methods.</p>