Ensemble pruning techniques are widely used to enhance a set of classifiers’ efficiency and predictive performance by selecting a subset of representative models, preventing redundancy, and ensuring diversity in classification tasks. The Optimum-Path Forest (OPF), a stable and efficient graph-based framework, offers versatile supervised and unsupervised capabilities in various machine-learning applications. The supervised version provides remarkable results with a simple graph-based structure produced by a training process conducted over a single dataset. However, one can notice little effort in OPF-based ensemble learning. This paper introduces an innovative approach to pruning OPF classifiers using meta-descriptions learned by Graph-Matching Networks, which are further employed to cluster similar OPF instances. The strategy selectively chooses representative models that excel in predictive tasks from groups generated by unsupervised OPF. Results demonstrate competitive performance to state-of-the-art pruning algorithms, with experiments conducted over fifteen public datasets, encouraging further exploration of Graph Matching Networks applied to ensemble pruning.