Geometric classification through ensemble learning

Image by Jametlene Reskp from Unsplash

Classifying complex data is one of the major challenges in machine learning, particularly when classes exhibit irregular geometries or highly non-linear decision boundaries. In these scenarios, graph-based methods provide a natural way to represent the underlying structure of the data and capture relationships that are often missed by traditional classifiers.

This research line focuses on the development of robust geometric classifiers that combine Minimum Spanning Trees (MSTs) with ensemble learning techniques. Unlike traditional approaches, which rely on a single spanning tree and are therefore highly sensitive to noise or small variations in the training data, these methods build multiple trees from different data subsets and exploit the consensus among them.

The information extracted from this ensemble is used to design new topological features that describe the degree to which each instance belongs to its class, providing a more stable and robust representation of the dataset. These features can then be used by different classification algorithms, decoupling structural analysis from the prediction process and combining the descriptive power of graph-based models with the flexibility of modern machine learning techniques. In (Aceña et al., 2026), we present a novel classification framework based on these ideas, introducing topological descriptors derived from ensembles of minimum spanning trees to improve robustness against noise and data variability. The proposed methodology is evaluated on a diverse collection of benchmark datasets, demonstrating competitive performance and highlighting the potential of ensemble-based graph representations for complex classification problems.