Dynamic Selection Best Model (DSBM) approach based on meta-learning

Ensemble learning methods such as bagging, boosting and stacking have emerged in recent decades as powerful machine learning techniques. They are based on the use of several learning models to improve prediction performance. However, they have certain drawbacks, particularly in terms of computational overhead, understanding, time execution and scalability. In this paper, we propose a new ensemble approach based on meta-learning that overcomes the limitations of existing ensemble methods. After training several classifiers, our method associates each instance with the most appropriate model, then builds a high-performance meta-model. Experimental validation results show that it outperforms the well-known bagging, boosting and stacking methods in terms of both accuracy and computation time. In addition, the results show that it is scalable.