Machine learning has been in the spotlight in the past decade, with astonishing results in many research domains. It is a consensus that deeply learnable features are a must-have, but improvements must be made on the classification layer. This talk discusses recent advances in the Optimum-Path Forest, a framework to design graph-based classification methods that cover supervised, unsupervised, and semi-supervised paradigms. Applications to deal with unbalanced datasets, recommender systems, and quantum computing will also be addressed.