Data complexity-based dynamic ensembling of SVMs in classification

Ensemble-based techniques are deployed to yield better performance than individual classifiers. Existing ensembling approaches fail to consider data complexity during their design. This work presents an ensemble-based approach for resolving complex patterns in real-world classification problems. A novel Minimum Spanning Tree (MST)-based approach is employed for decomposing the original problem into subproblems with reduced data complexity, and SVM is utilized for the development of component classifiers for these subproblems. A novel dynamic ensemble-based technique is utilized to generate the outcome. This work is evaluated by using 28 datasets retrieved from the KEEL dataset repository. Additionally, statistical tests are performed to illustrate a significant difference in the performance of the proposed model compared to the state-of-art classification models and two recently proposed dynamic ensembling approaches. •Novel Ensemble-based technique.•Divides the original problem into subproblems of lower data complexity.•Utilizes Minimum Spanning Tree for creating subproblems.•Creates models for subproblems using SVM.•Novel approach for dynamic selection of the SVM model to yield the final outcome.