The growing demand for privacy-preserving machine learning has positioned federated learning as a promising research paradigm, enabling the training of high-performance models across distributed data sources without compromising user privacy. However, despite its advantages, federated learning faces critical challenges arising from the heterogeneity and volatility of participating clients. In real-world scenarios, variations in client participation, data volume, computational capability, and communication reliability contribute to a highly dynamic training environment, which negatively impacts efficiency and convergence of the model. To address these challenges, this paper proposes a novel client selection method named CDE3. First, CDE3 employs a multidimensional model to comprehensively evaluate clients’ contributions. Second, we enhance the classical Exp3 algorithm by incorporating a discount factor that exponentially decays historical contributions, thereby increasing the influence of recent client behavior in the selection process. Furthermore, we provide a theoretical analysis demonstrating a favorable regret bound for the proposed method. Extensive experiments conducted in volatile FL settings validate the effectiveness of CDE3, showing improved convergence speed and model accuracy compared with those of the baseline algorithms. These results confirm that CDE3 effectively mitigates volatility, enhancing the stability and efficiency of federated learning.