Abstarct: Customer churn is a significant issue for banks and financial institutions, requiring continuous analysis of factors such as resources, costs, and effective interest rates. The aim of this research is to automatically and early detect customers at risk of churn using machine learning techniques. One of the challenges in this subject is the data imbalance due to the low proportion of churned customers. This research seeks to investigate the role of data quality on the effectiveness of results through ensemble learning, without disrupting data distribution and with minimal addition of synthetic data. To handle imbalanced data, methods such as clustering and a combination of over-sampling and under-sampling (SMOTEENN method) have been used. These techniques are applied to both majority and minority classes, then evaluated using Support Vector Machines and weighted majority voting ensemble learning, along with hyperparameter optimization through the Imperialist Competitive Algorithm (ICA). The results indicate that the combination of these methods, especially in Recall and F1-score for minority class data, performs better than baxseline and advanced models such as Random Forest and Multi-laxyer Perceptron (MLP) in imbalanced conditions. This research highlights the importance of using ensemble methods, the impact of data quality, and the selection of balancing models in improving the accuracy of customer churn prediction systems.