Ensemble learning encompasses methods that generate many well-diversified predic-tors and aggregates their results to perform a better prediction. These predictors are usually weak and low-cost for obtaining when they are alone. However, they reveal excellent per-formance when they are skillfully used together in the form of a learning architecture. Metaheuristic methods have been used to form such architecture optimally during recent years. Along this stream, in this paper, a bi-level optimization based on discrete-continuous genetic algorithm is utilized to enhance the performance of an ensemble learning meta-algorithm which benefits decision tree classification. Feature selection and tree model con-structing for any ensemble member are done by the metaheuristic method. It allows us to have advantages of tree-based prediction models, ensemble learning, and solution optimal-ity simultaneously. The proposed system is compared to some well-known ensemble learn-ing methods. Results show significant superiority of the proposed system in terms of pre-diction accuracy.